{ "exportedAt": "2026-10-04T17:37:27.491Z", "papers": [ { "id": "arxiv-2610.00854", "title": "Are Frontier VLM Agents Ready to Be Robot Generalists? An Empirical Study with the Embodied Agent Arena", "titleZh": "前沿视觉语言模型智能体准备好成为通用机器人了吗？基于具身智能体竞技场的实证研究", "date": "2026-10-01", "year": 2026, "url": "https://arxiv.org/abs/2610.00854", "paperUrl": "https://arxiv.org/abs/2610.00854", "projectUrl": null, "codeUrl": null, "category": "评测与基准", "tags": [ "VLM智能体", "具身评测", "几何推理", "任务完成" ], "directions": [ "数据集与基准" ], "robotFilters": [], "tier": "recent", "summary": "跨五种能力的1000案例评测区分局部感知能力与完整机器人目标达成。", "abstractZh": "Embodied Agent Arena把几何、空间推理、可供性、规划与操作放入保留各来源接口的统一测试框架。七种VLM中Astra总体较强，但更丰富视觉反馈会改变控制排名，局部精度和接触定位仍不能保证完整目标完成。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "比较7种VLM；固定输入每例通常5次，Fable为3次，交互和轨迹采用单次记录。操纵基线含182个二元任务；153共同任务比较基线与逐轮RGB扩展，并用46共同案例检验自审、裁剪审查及工具辅助。", "robots": [ "CALVIN、RLBench、RoboTwin等基准机器人（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现应锁定模型版本、各来源预算、观测协议和成功分母，缺失记录与无效连续估计需分开处理。重点复查终态条件、动作调用和共同任务子集，不能将不同RGB配置或单次交互波动混成稳定模型排名。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "整合32个既有来源与新GeoProbe，共1000案例，保留原始观测、动作接口和成功条件。固定输入任务使用持久Python交互，操作遵循原生控制循环；分别计几何误差、接触有效性、规划Pass@1及终态所有目标条件，避免用局部进展替代成功。", "whyUseful": "复现应锁定模型版本、各来源预算、观测协议和成功分母，缺失记录与无效连续估计需分开处理。重点复查终态条件、动作调用和共同任务子集，不能将不同RGB配置或单次交互波动混成稳定模型排名。", "limitations": "并非实体机器人部署，真实照片只用于感知测试；物理迁移仍待验证。公开资产、来源异质辅助工具和不等重试次数限制泛化解释，交互任务重复不足；合成跨域分数不能替代各任务原生指标。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00854v1", "note": "§3–4.1，PDF页3–6：1000案例、32来源加GeoProbe；固定输入多次但交互任务单次。" }, { "url": "https://arxiv.org/pdf/2610.00854v1", "note": "表2–3，PDF页6、10：操作41.8%为基线；RGB扩展共同153任务产生明显排名变化。" }, { "url": "https://arxiv.org/pdf/2610.00854v1", "note": "§5，PDF页11：明确物理机器人迁移与长期错误恢复仍须后续评估。" }, { "url": "https://arxiv.org/pdf/2610.00854v1", "note": "PDF页10（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823236+00:00", "original": { "id": "arxiv-2610.00854", "originalSourceUrl": "https://arxiv.org/pdf/2610.00854v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00854v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00854", "sha256": "b824475f46cfe7a1e9a67263286e751e2309101294a5e8c95425c5888f80fe77", "bytes": 21941418, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:24.559531+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00854", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00854", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00854", "pages": 40, "sourceIdentityStatus": "title_match", "sourceTitle": "Are Frontier VLM Agents Ready to Be Robot Generalists? An Empirical Study with the Embodied Agent Arena", "sourceVersion": "2610.00854v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.633904+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00854" }, "originalAnalysis": { "id": "arxiv-2610.00854", "sectionsRead": [ "§3，PDF页3–5", "§4.1–4.3及表2，PDF页6–8", "§4.4及表3，PDF页10–11", "§5限制，PDF页11" ], "analyzedAt": "2026-10-04T13:58:28.823236+00:00", "analysisStatus": "full_text_sections", "methodsZh": "整合32个既有来源与新GeoProbe，共1000案例，保留原始观测、动作接口和成功条件。固定输入任务使用持久Python交互，操作遵循原生控制循环；分别计几何误差、接触有效性、规划Pass@1及终态所有目标条件，避免用局部进展替代成功。", "experimentsZh": "比较7种VLM；固定输入每例通常5次，Fable为3次，交互和轨迹采用单次记录。操纵基线含182个二元任务；153共同任务比较基线与逐轮RGB扩展，并用46共同案例检验自审、裁剪审查及工具辅助。", "resultsZh": "Astra空间Pass@1为69.4%，规划80.3%，操纵41.8%；位姿误差17.1度/106.6厘米。RGB扩展后共同任务Astra仍63/153成功，Fable从34升至65；说明总分可隐藏相反任务效应。", "limitationsZh": "并非实体机器人部署，真实照片只用于感知测试；物理迁移仍待验证。公开资产、来源异质辅助工具和不等重试次数限制泛化解释，交互任务重复不足；合成跨域分数不能替代各任务原生指标。", "reproductionZh": "复现应锁定模型版本、各来源预算、观测协议和成功分母，缺失记录与无效连续估计需分开处理。重点复查终态条件、动作调用和共同任务子集，不能将不同RGB配置或单次交互波动混成稳定模型排名。", "sourceUrl": "https://arxiv.org/pdf/2610.00854v1", "sourceVersion": "2610.00854v1", "originalSha256": "b824475f46cfe7a1e9a67263286e751e2309101294a5e8c95425c5888f80fe77", "experimentType": "sim", "robots": [ "CALVIN、RLBench、RoboTwin等基准机器人（仿真）" ], "evidenceNotes": [ { "section": "§3–4.1，PDF页3–6", "note": "1000案例、32来源加GeoProbe；固定输入多次但交互任务单次。" }, { "section": "表2–3，PDF页6、10", "note": "操作41.8%为基线；RGB扩展共同153任务产生明显排名变化。" }, { "section": "§5，PDF页11", "note": "明确物理机器人迁移与长期错误恢复仍须后续评估。" }, { "section": "PDF页10（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "机器人通用能力诊断", "执行协议敏感性" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Frontier vision-language models (VLMs) combine scene estimation, interaction grounding, and executable actions. Understanding how these abilities support complete robotic tasks is central to evaluating their readiness as robot generalists. We introduce Embodied Agent Arena to examine where local competence supports, or falls short of, complete task success across Geometry, Spatial Reasoning, Affordance, Task Planning, and Manipulation. The arena contains 1,000 cases drawn from 32 established sources and GeoProbe, our new benchmark for geometric estimation on Blender renders and real-scene images. A minimal harness preserves source observations and operations while separating metric precision, functional grounding, and native goal completion. We evaluate seven VLMs, analyze Astra's task-specific advantages, and compare richer-observation execution protocols and multi-round review. Across the arena, Astra's advantage is strongest in precise estimation and usable-contact localization; completing coordinated, goal-directed actions remains the key gap to robot generalism." }, { "id": "arxiv-2610.00864", "title": "Kinematic MeanFlow: One-Step Action Generation Policy for Robotic Foundation Models", "titleZh": "运动学MeanFlow：用于机器人基础模型的单步动作生成策略", "date": "2026-10-01", "year": 2026, "url": "https://arxiv.org/abs/2610.00864", "paperUrl": "https://arxiv.org/abs/2610.00864", "projectUrl": null, "codeUrl": null, "category": "视觉语言动作模型", "tags": [ "单步生成", "流匹配加速", "MeanFlow", "推理效率" ], "directions": [ "视觉语言动作" ], "robotFilters": [ "GoogleX", "WidowX", "Unitree G1" ], "tier": "recent", "summary": "拆分速度时间导数，稳定机器人基础模型的单步动作生成。", "abstractZh": "K-MF针对MeanFlow直接用于机器人动作时的训练崩溃，将时间导数分解到两个去噪子区间。实验覆盖基础模型微调和从头训练，性能与多步流匹配相当或更好；全部机器人策略评测均为仿真，另测GPU推理速度。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "GR00T-N1.6微调及SimVLA-S从头训练，在LIBERO、COMPASS PointNav及SimplerEnv评估。Fractal/BridgeData V2虽来自实机，本文属于Real2Sim；导航Unitree G1也在仿真。另在L40和Jetson Orin测eager/compile延迟。", "robots": [ "GoogleX（SimplerEnv仿真）", "WidowX（SimplerEnv仿真）", "Unitree G1（COMPASS仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现须加入第二时间嵌入、正确stop-gradient与JVP切向量，区分训练迭代数和NFE。原文提供超参与算法，基线MeanFlow需更精细训练；应同时报告等迭代训练成本、动作头与整模型延迟，代码实际开放状态未核验。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "依据区间平均速度恒等式，将MeanFlow目标中的总时间导数拆成早期、晚期两部分，缓解末期局部加速度突增和样本间尺度扩张。用前向自动微分JVP并行计算两项，比较随机、中点和小MLP学习分割点；推理一次网络评估直接输出动作块。", "whyUseful": "复现须加入第二时间嵌入、正确stop-gradient与JVP切向量，区分训练迭代数和NFE。原文提供超参与算法，基线MeanFlow需更精细训练；应同时报告等迭代训练成本、动作头与整模型延迟，代码实际开放状态未核验。", "limitations": "未提供实机闭环验证，未覆盖全身控制或灵巧操作；训练时间较流匹配增加约27–33%。性能并非逐任务全部提升，例如Bridge胡萝卜和Fractal关门略降；硬件延迟收益依赖主干占比及编译设置。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00864v1", "note": "§4.1，PDF页6；表3，PDF页8：明确Sim2Sim与Real2Sim，不能将真实训练数据当实机实验。" }, { "url": "https://arxiv.org/pdf/2610.00864v1", "note": "表2，PDF页7；表5，PDF页9：成功率与L40/Orin延迟来自不同实验设置。" }, { "url": "https://arxiv.org/pdf/2610.00864v1", "note": "PDF页9（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823234+00:00", "original": { "id": "arxiv-2610.00864", "originalSourceUrl": "https://arxiv.org/pdf/2610.00864v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00864v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00864", "sha256": "8b4929f164860d768f2960625bb5a3a9b2b0ae7e577fa0e0a94106628f14f84c", "bytes": 22731291, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:24.514221+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00864", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00864", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00864", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "Kinematic MeanFlow: One-Step Action Generation Policy for Robotic Foundation Models", "sourceVersion": "2610.00864v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.611285+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00864" }, "originalAnalysis": { "id": "arxiv-2610.00864", "sectionsRead": [ "§3，PDF页4–6", "§4.1–4.2及表1–2，PDF页6–7", "§4.3–4.5及表3–5，PDF页8–9", "§5限制，PDF页9" ], "analyzedAt": "2026-10-04T13:58:28.823234+00:00", "analysisStatus": "full_text_sections", "methodsZh": "依据区间平均速度恒等式，将MeanFlow目标中的总时间导数拆成早期、晚期两部分，缓解末期局部加速度突增和样本间尺度扩张。用前向自动微分JVP并行计算两项，比较随机、中点和小MLP学习分割点；推理一次网络评估直接输出动作块。", "experimentsZh": "GR00T-N1.6微调及SimVLA-S从头训练，在LIBERO、COMPASS PointNav及SimplerEnv评估。Fractal/BridgeData V2虽来自实机，本文属于Real2Sim；导航Unitree G1也在仿真。另在L40和Jetson Orin测eager/compile延迟。", "resultsZh": "LIBERO平均成功率：GR00T单步97.9%对四步97.3%，SimVLA单步95.1%对十步94.2%。Fractal、Bridge评测为78.4%/59.9%。动作头延迟降低67.5–74.4%，端到端降低30.3–54.9%。", "limitationsZh": "未提供实机闭环验证，未覆盖全身控制或灵巧操作；训练时间较流匹配增加约27–33%。性能并非逐任务全部提升，例如Bridge胡萝卜和Fractal关门略降；硬件延迟收益依赖主干占比及编译设置。", "reproductionZh": "复现须加入第二时间嵌入、正确stop-gradient与JVP切向量，区分训练迭代数和NFE。原文提供超参与算法，基线MeanFlow需更精细训练；应同时报告等迭代训练成本、动作头与整模型延迟，代码实际开放状态未核验。", "sourceUrl": "https://arxiv.org/pdf/2610.00864v1", "sourceVersion": "2610.00864v1", "originalSha256": "8b4929f164860d768f2960625bb5a3a9b2b0ae7e577fa0e0a94106628f14f84c", "experimentType": "sim", "robots": [ "GoogleX（SimplerEnv仿真）", "WidowX（SimplerEnv仿真）", "Unitree G1（COMPASS仿真）" ], "evidenceNotes": [ { "section": "§4.1，PDF页6；表3，PDF页8", "note": "明确Sim2Sim与Real2Sim，不能将真实训练数据当实机实验。" }, { "section": "表2，PDF页7；表5，PDF页9", "note": "成功率与L40/Orin延迟来自不同实验设置。" }, { "section": "PDF页9（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "动作生成加速", "机器人基础模型训练" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "In this paper, we study how to achieve one-step action generation in Robotic Foundation Models (RFMs), aiming to overcome the high inference latency of multi-step flow matching. MeanFlow provides a promising framework for this goal, yet its direct application leads to performance collapse. We discover that this stems from two distinctive dynamics exhibited in the RFM velocity field: (1) the ``local acceleration\" exhibits stability early on, but surges sharply towards the end of the denoising process, and (2) the spread of its magnitudes across samples widens as denoising progresses. To address these issues, we introduce Kinematic MeanFlow (K-MF), a novel one-step action policy tailored for RFMs. Specifically, grounded in a kinematic identity, K-MF decouples the time derivative term in the MeanFlow formulation into two sub-interval terms separated by an intermediate point. This decoupled formulation enables the two terms to capture early-stage and late-stage denoising dynamics, respectively, while mitigating the error amplification across the process. As a result, our K-MF empowers RFMs to achieve one-step action generation in both training from scratch and fine-tuning paradigms across diverse tasks, while outperforming multi-step flow matching in most settings. In terms of inference efficiency, K-MF reduces action-head latency of GR00T-N1.6 by 67.5%~74.4% across L40 and Jetson Orin in eager and compiled modes, yielding end-to-end latency reductions of 30.3%~54.9%. Code will be available at https://github.com/IntelChina-AI/K-MF." }, { "id": "arxiv-2610.00878", "title": "UniTrackPLA: Unified Panorama-Language-Action Model for Instruction-Guided Navigation and Dynamic Person Tracking", "titleZh": "UniTrackPLA：用于指令引导导航与动态人员跟踪的统一全景—语言—动作模型", "date": "2026-10-01", "year": 2026, "url": "https://arxiv.org/abs/2610.00878", "paperUrl": "https://arxiv.org/abs/2610.00878", "projectUrl": null, "codeUrl": null, "category": "导航与具身智能", "tags": [ "全景感知", "人员跟随", "视觉语言导航", "世界动作一致性" ], "directions": [ "视觉语言动作", "导航与建图" ], "robotFilters": [ "Unitree Go2-W" ], "tier": "recent", "summary": "统一全景导航与人员跟随，并用未来特征一致性决定复用动作或重规划。", "abstractZh": "模型把360度全景分解为保留方位和时间身份的透视视图，输出机器人坐标系航点块；世界—动作一致性模块检查执行后观测是否符合预测。论文包含仿真评测、手持相机真实路线数据及Go2-W实机定性部署。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "both", "experimentNote": "数据含5000条仿真跟随、10000条仿真导航和96条手持全景真实路线；仿真测试200跟随回合与400导航路线。真实数据76条训练、20条留出，另将模型部署至带Insta360 X4的Go2-W开展室内外跟随与导航。", "robots": [ "Unitree Go2-W（实机）", "OmniTrackNav-Bench导航代理（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现需一致全景投影、方位/时间顺序、按路线隔离的数据划分及验证集选阈值。原文用8张RTX3090训练1轮、4个180度重叠视图和最多31帧历史；应分别复现离线航点误差、仿真闭环和实体案例。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "PAE将全景投为循环透视视图，冻结DINOv3、SigLIP和Qwen3主干，融合方位和历史时间编码。动作头预测8个累计航点；WAC预测前4步特征，按余弦差阈值决定继续或重新规划，辅助梯度只更新WAC。", "whyUseful": "复现需一致全景投影、方位/时间顺序、按路线隔离的数据划分及验证集选阈值。原文用8张RTX3090训练1轮、4个180度重叠视图和最多31帧历史；应分别复现离线航点误差、仿真闭环和实体案例。", "limitations": "96条路线本身是手持相机数据，不是96次自主机器人完成；实机部分主要是定性演示，未给同规模闭环成功率。整体仿真导航成功仍低，拥挤跟随尤其困难；WiFi远端推理限制完全机载使用。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00878v1", "note": "§IV-A，PDF页5–6：真实96条路线9.13公里来自手持全景相机；与实机执行证据区分。" }, { "url": "https://arxiv.org/pdf/2610.00878v1", "note": "表I–III，PDF页6：跟随/导航成功率来自仿真，真实留出指标为航点预测误差。" }, { "url": "https://arxiv.org/pdf/2610.00878v1", "note": "§IV-E与图5–9，PDF页7–8：Go2-W实体部署及远端RTX4090推理，有实机定性证据。" }, { "url": "https://arxiv.org/pdf/2610.00878v1", "note": "PDF页7（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823229+00:00", "original": { "id": "arxiv-2610.00878", "originalSourceUrl": "https://arxiv.org/pdf/2610.00878v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00878v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00878", "sha256": "96d03780cfb2ae8e92a1e3aa4456520845d5d17ab48fcce1ed469e403c8b4706", "bytes": 14192716, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:15.091145+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00878", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00878", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00878", "pages": 9, "sourceIdentityStatus": "title_match", "sourceTitle": "UniTrackPLA: Unified Panorama-Language-Action Model for Instruction-Guided Navigation and Dynamic Person Tracking", "sourceVersion": "2610.00878v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.592815+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00878" }, "originalAnalysis": { "id": "arxiv-2610.00878", "sectionsRead": [ "§III，PDF页3–5", "§IV-A–C及表I–III，PDF页5–6", "§IV-D–E及图5–9，PDF页7–8", "§V，PDF页8" ], "analyzedAt": "2026-10-04T13:58:28.823229+00:00", "analysisStatus": "full_text_sections", "methodsZh": "PAE将全景投为循环透视视图，冻结DINOv3、SigLIP和Qwen3主干，融合方位和历史时间编码。动作头预测8个累计航点；WAC预测前4步特征，按余弦差阈值决定继续或重新规划，辅助梯度只更新WAC。", "experimentsZh": "数据含5000条仿真跟随、10000条仿真导航和96条手持全景真实路线；仿真测试200跟随回合与400导航路线。真实数据76条训练、20条留出，另将模型部署至带Insta360 X4的Go2-W开展室内外跟随与导航。", "resultsZh": "仿真跟随成功率35%对复现TrackVLA的23.5%；导航19.75%对13%。加入76条真实路线使留出ADE从0.172降至0.038米。实机通过远端RTX4090约6.9 FPS推理、3.5Hz执行，展示多场景案例。", "limitationsZh": "96条路线本身是手持相机数据，不是96次自主机器人完成；实机部分主要是定性演示，未给同规模闭环成功率。整体仿真导航成功仍低，拥挤跟随尤其困难；WiFi远端推理限制完全机载使用。", "reproductionZh": "复现需一致全景投影、方位/时间顺序、按路线隔离的数据划分及验证集选阈值。原文用8张RTX3090训练1轮、4个180度重叠视图和最多31帧历史；应分别复现离线航点误差、仿真闭环和实体案例。", "sourceUrl": "https://arxiv.org/pdf/2610.00878v1", "sourceVersion": "2610.00878v1", "originalSha256": "96d03780cfb2ae8e92a1e3aa4456520845d5d17ab48fcce1ed469e403c8b4706", "experimentType": "both", "robots": [ "Unitree Go2-W（实机）", "OmniTrackNav-Bench导航代理（仿真）" ], "evidenceNotes": [ { "section": "§IV-A，PDF页5–6", "note": "真实96条路线9.13公里来自手持全景相机；与实机执行证据区分。" }, { "section": "表I–III，PDF页6", "note": "跟随/导航成功率来自仿真，真实留出指标为航点预测误差。" }, { "section": "§IV-E与图5–9，PDF页7–8", "note": "Go2-W实体部署及远端RTX4090推理，有实机定性证据。" }, { "section": "PDF页7（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "统一导航策略", "全景闭环控制" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "General-purpose embodied robots should support both navigation toward language-specified destinations and dynamic person tracking under arbitrary initial target azimuths. However, existing methods typically rely on forward-facing observations and address these tasks with separate policies, limiting omnidirectional perception and unified closed-loop control. We present UniTrackPLA, a unified panorama-language-action model for instruction-guided navigation and dynamic person tracking. Its Panoramic-Aware Encoding (PAE) preserves the temporal and azimuthal structure of perspective views projected from each panorama, enabling perspective-pretrained visual encoders to process omnidirectional observations. A shared vision-language backbone grounds instructions in the panoramic context and predicts continuous robot-centric waypoint chunks for both tasks. World-Action Consistency (WAC) further predicts action-conditioned future visual states and verifies waypoint prefixes online, allowing reliable actions to be reused while triggering replanning upon inconsistency. We also introduce OmniTrackNav-Bench, comprising 5,000 simulated tracking trajectories, 10,000 simulated VLN routes, and 96 verified real-world routes, providing 919,978 waypoint-supervision instances. UniTrackPLA improves overall tracking SR from 23.50% to 35.00% and Omni-VLN SR/SPL from 13.00%/12.77% to 19.75%/19.29%. Incorporating 76 real-world routes further improves held-out EP@0.2m from 42.92% to 92.08%. Closed-loop experiments on a Go2-W robot demonstrate unified panoramic tracking and navigation across indoor and outdoor environments. The project page is at https://tw5775.github.io/UniTrackPLA." }, { "id": "arxiv-2610.00897", "title": "Real-Time Human-Adaptive Task Allocation for Multi-Human Multi-Robot Supervision", "titleZh": "面向多人多机器人监督的实时人因自适应任务分配", "date": "2026-10-01", "year": 2026, "url": "https://arxiv.org/abs/2610.00897", "paperUrl": "https://arxiv.org/abs/2610.00897", "projectUrl": null, "codeUrl": null, "category": "人机协作与多机器人", "tags": [ "人因", "任务分配", "认知负荷", "用户研究" ], "directions": [], "robotFilters": [], "tier": "recent", "summary": "依据操作员表现、眼动和主观状态动态分配监督任务，提高仿真仓储团队吞吐。", "abstractZh": "HAMA以可调整的操作员容量替代固定监督上限，在每个实验时段后更新容量，并在线将高优先级请求分给预计相对负荷最低者。30人参与的VR仿真研究支持任务完成和眨眼指标改善，但主观工作负荷、疲劳差异不显著。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "30名参与者组成10个三人团队，在Unity仓库监督20台模拟移动机器人，处理死锁和手动充电停靠。被试内对比固定容量基线、平衡顺序；每条件10个4分钟时段，主要统计第3–10时段。", "robots": [ "Unity仓储移动机器人（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现需Meta Quest Pro眼动校准、VR多窗口与键盘控制、2-back及每时段问卷。应共同控制边界重分配，在基线仅关闭容量与注意力权重，记录团队相关性，并分别报告客观表现、眨眼和主观量表。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "初始容量来自三次2-back测试；随后综合NASA-TLX、疲劳评分、任务完成数和眨眼变化，按时段增减一个监督名额。以视线停留时间估计不同任务的相对难度，贪心分配最紧急任务，并按死锁簇边界优先调度可移动机器人。", "whyUseful": "复现需Meta Quest Pro眼动校准、VR多窗口与键盘控制、2-back及每时段问卷。应共同控制边界重分配，在基线仅关闭容量与注意力权重，记录团队相关性，并分别报告客观表现、眨眼和主观量表。", "limitations": "真人用户研究不代表实体机器人试验，机器人运行于Unity。疲劳代理指标不能当临床或主观疲劳改善证据；样本以年轻、缺乏系统经验者为主，仅两类干预，且容量在时段边界而非持续更新。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00897v1", "note": "§V，PDF页5–6：30人、10团队、Unity仿真、20移动机器人；并非实机仓库部署。" }, { "url": "https://arxiv.org/pdf/2610.00897v1", "note": "表III及§VI，PDF页7–8：任务完成和眨眼显著；NASA-TLX p=.975、主观疲劳p=.224。" }, { "url": "https://arxiv.org/pdf/2610.00897v1", "note": "PDF页7（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823227+00:00", "original": { "id": "arxiv-2610.00897", "originalSourceUrl": "https://arxiv.org/pdf/2610.00897v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00897v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00897", "sha256": "f3ce897e305834e7e2343caac8438cb3e4e3e7589ac3a9d2a53761ade95076c0", "bytes": 5265792, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:14.872588+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00897", "license": "http://creativecommons.org/licenses/by-nc-nd/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00897", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00897", "pages": 10, "sourceIdentityStatus": "title_match", "sourceTitle": "Real-Time Human-Adaptive Task Allocation for Multi-Human Multi-Robot Supervision", "sourceVersion": "2610.00897v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.579733+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00897" }, "originalAnalysis": { "id": "arxiv-2610.00897", "sectionsRead": [ "§IV，PDF页4–5", "§V用户研究，PDF页5–6", "§VI及表II–III，PDF页7–8", "§VII-C限制，PDF页8" ], "analyzedAt": "2026-10-04T13:58:28.823227+00:00", "analysisStatus": "full_text_sections", "methodsZh": "初始容量来自三次2-back测试；随后综合NASA-TLX、疲劳评分、任务完成数和眨眼变化，按时段增减一个监督名额。以视线停留时间估计不同任务的相对难度，贪心分配最紧急任务，并按死锁簇边界优先调度可移动机器人。", "experimentsZh": "30名参与者组成10个三人团队，在Unity仓库监督20台模拟移动机器人，处理死锁和手动充电停靠。被试内对比固定容量基线、平衡顺序；每条件10个4分钟时段，主要统计第3–10时段。", "resultsZh": "平均完成数17.86对16.22，约增10%，算法效应p=0.046；眨眼55.92对66.39，约降16%，p=0.020。NASA-TLX几乎不变，疲劳5.15对5.54但不显著；平均分配延迟1.34毫秒。", "limitationsZh": "真人用户研究不代表实体机器人试验，机器人运行于Unity。疲劳代理指标不能当临床或主观疲劳改善证据；样本以年轻、缺乏系统经验者为主，仅两类干预，且容量在时段边界而非持续更新。", "reproductionZh": "复现需Meta Quest Pro眼动校准、VR多窗口与键盘控制、2-back及每时段问卷。应共同控制边界重分配，在基线仅关闭容量与注意力权重，记录团队相关性，并分别报告客观表现、眨眼和主观量表。", "sourceUrl": "https://arxiv.org/pdf/2610.00897v1", "sourceVersion": "2610.00897v1", "originalSha256": "f3ce897e305834e7e2343caac8438cb3e4e3e7589ac3a9d2a53761ade95076c0", "experimentType": "sim", "robots": [ "Unity仓储移动机器人（仿真）" ], "evidenceNotes": [ { "section": "§V，PDF页5–6", "note": "30人、10团队、Unity仿真、20移动机器人；并非实机仓库部署。" }, { "section": "表III及§VI，PDF页7–8", "note": "任务完成和眨眼显著；NASA-TLX p=.975、主观疲劳p=.224。" }, { "section": "PDF页7（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "监督容量自适应", "人机团队协同" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "We propose a human-factor-aware method of allocating robot supervision tasks to multiple human operators. In scenarios where multiple operators occasionally teleoperate multiple robots to help the robots overcome difficulties, the allocation of the supervisory control tasks to humans needs to consider the real-time cognitive states of individual operators. However, most existing methods assume fixed supervisory capacity per operator and overlook fluctuations in the human factors such as workload and fatigue. As a result, workload distribution can be unbalanced where some operators become overloaded while the others remain underused. Our method dynamically regulates supervisory capacity and allocates tasks in a way that maintains balanced mental workload, prevents overload, and improves overall team performance. The allocation method uses a greedy strategy that minimizes estimated operator workloads with task prioritization. Robots are assigned to operators by reflecting their current supervisory capacity where the required effort depends on the types of tasks. In the user study, the analysis across predefined time intervals shows that the proposed method consistently achieves higher performance and lower behavioral signs of fatigue compared to a baseline method that does not consider human factors. These results highlight adaptive capacity adjustment as an effective preventive mechanism for sustaining operator performance in long-duration, high-demand settings." }, { "id": "arxiv-2610.00899", "title": "TOAST: Stochastic Robot Action Tokenization for Autoregressive Vision-Language-Action Models", "shortTitle": "TOAST", "date": "2026-10-01", "category": "VLA / 动作表示", "freshness": "最新预印本", "summary": "先对动作块做DCT、量化及按维度展开，再用unigram词表枚举等价分词；训练时从64个候选中采样，而各候选解码后的量化动作完全相同。它改变交叉熵监督的分词边界，不增加示范，也不改变策略架构。", "whyUseful": "优先复现等价解码单测及1/8、1/16数据消融；固定512词表、DCT缩放10、训练3万步，并严格区分外部DROID词表与目标数据词表。实机为20Hz关节动作及1秒开环动作块。", "paperUrl": "https://arxiv.org/abs/2610.00899", "projectUrl": "https://kskshr.github.io/toast/", "codeUrl": null, "trainingStatus": "待开源", "trainingNote": "作者项目页明确标注 Code (Coming Soon)。", "caveats": "2026-10-01 新预印本；数值为作者报告，尚未独立复现。", "url": "https://arxiv.org/abs/2610.00899", "status": "待开源", "titleZh": "TOAST：面向自回归视觉—语言—动作模型的随机机器人动作词元化", "abstractZh": "先对动作块做DCT、量化及按维度展开，再用unigram词表枚举等价分词；训练时从64个候选中采样，而各候选解码后的量化动作完全相同。它改变交叉熵监督的分词边界，不增加示范，也不改变策略架构。\n同一PaliGemma-3B管线比较LIBERO五档数据量，并在Franka Research 3上做收桌、装袋、早餐摆放和抽屉收纳；各方法共用示范子集，三训练种子，仿真每种子2000回合，实机每任务每种子30回合。", "experimentType": "both", "experimentNote": "同一PaliGemma-3B管线比较LIBERO五档数据量，并在Franka Research 3上做收桌、装袋、早餐摆放和抽屉收纳；各方法共用示范子集，三训练种子，仿真每种子2000回合，实机每任务每种子30回合。", "robots": [ "Franka Emika Panda（仿真）", "Franka Research 3（实机）" ], "robotNote": "Panda 用于 LIBERO 仿真；Research 3 配 Robotiq 2F-85 夹爪用于实机。", "tags": [ "VLA", "动作词元化", "少样本学习", "模仿学习", "单臂操作" ], "contribution": "先对动作块做DCT、量化及按维度展开，再用unigram词表枚举等价分词；训练时从64个候选中采样，而各候选解码后的量化动作完全相同。它改变交叉熵监督的分词边界，不增加示范，也不改变策略架构。", "limitations": "仅研究Franka形态；不直接适用学习式向量量化，未比较BPE-dropout，词表大小与候选数固定。实机收益不能只归因于示范较少，因为动作表示和任务也不同。", "codeStatus": "pending", "codeStatusNote": "沿用原清单已核对的项目页 Code (Coming Soon)；本轮项目页无法重读，正文再次确认官方项目地址。", "license": "未核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/html/2610.00899v1", "note": "摘要及 §V-A、§V-C 明确仿真/实机、Panda/Research 3 型号；§VI 给出适用边界。" }, { "url": "https://kskshr.github.io/toast/", "note": "官方地址由论文给出；待开源状态来自原清单的项目页核对，本轮页面读取失败。" }, { "url": "https://arxiv.org/html/2610.00899v1", "note": "IV; V-A：随机等价分词、统一训练设置及硬件。" }, { "url": "https://arxiv.org/html/2610.00899v1", "note": "Table III; Fig.5; VI：全量非最优、低数据与实机增益及明确限制。" }, { "url": "https://arxiv.org/pdf/2610.00899v1", "note": "IV; V-A：随机等价分词、统一训练设置及硬件。" }, { "url": "https://arxiv.org/pdf/2610.00899v1", "note": "Table III; Fig.5; VI：全量非最优、低数据与实机增益及明确限制。" } ], "tier": "recent", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对范围为公开原文、平台型号与所列来源；性能数字均为作者报告，未独立复现。", "year": 2026, "directions": [ "视觉语言动作", "操作与抓取", "模仿学习" ], "robotFilters": [ "Franka Panda", "Franka Research 3" ], "original": { "id": "arxiv-2610.00899", "originalSourceUrl": "https://arxiv.org/pdf/2610.00899v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00899v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00899", "sha256": "7d32b2fe30393f482558eb50947f5764c554703fa7012d91821ba92e7d6627ef", "bytes": 5374014, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:33:10.604532+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00899", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00899", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00899", "pages": 10, "sourceIdentityStatus": "title_match", "sourceTitle": "TOAST: Stochastic Robot Action Tokenization for Autoregressive Vision-Language-Action Models", "sourceVersion": "2610.00899v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.140170+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00899" }, "originalAnalysis": { "id": "arxiv-2610.00899", "sourceUrl": "https://arxiv.org/pdf/2610.00899v1", "sectionsRead": [ "III", "IV-A–C", "V-A–D", "VI", "Tables I–IV; Figs 4–8" ], "methodsZh": "先对动作块做DCT、量化及按维度展开，再用unigram词表枚举等价分词；训练时从64个候选中采样，而各候选解码后的量化动作完全相同。它改变交叉熵监督的分词边界，不增加示范，也不改变策略架构。", "experimentsZh": "同一PaliGemma-3B管线比较LIBERO五档数据量，并在Franka Research 3上做收桌、装袋、早餐摆放和抽屉收纳；各方法共用示范子集，三训练种子，仿真每种子2000回合，实机每任务每种子30回合。", "resultsZh": "LIBERO全量91.4%，并未超过重建FAST的92.4%；1/16数据时42.8%，比确定性自身高6.8个百分点。四项实机平均50.8%，提高15.8个百分点；分层bootstrap支持实机及低数据差异。采样开销约0.4%。", "limitationsZh": "仅研究Franka形态；不直接适用学习式向量量化，未比较BPE-dropout，词表大小与候选数固定。实机收益不能只归因于示范较少，因为动作表示和任务也不同。", "reproductionZh": "优先复现等价解码单测及1/8、1/16数据消融；固定512词表、DCT缩放10、训练3万步，并严格区分外部DROID词表与目标数据词表。实机为20Hz关节动作及1秒开环动作块。", "experimentType": "both", "robots": [ "Franka Emika Panda（仿真）", "Franka Research 3（实机）" ], "evidenceNotes": [ { "note": "随机等价分词、统一训练设置及硬件。", "section": "IV; V-A" }, { "note": "全量非最优、低数据与实机增益及明确限制。", "section": "Table III; Fig.5; VI" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "corrections": [], "sourceVersion": "2610.00899v1", "originalSha256": "7d32b2fe30393f482558eb50947f5764c554703fa7012d91821ba92e7d6627ef" }, "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2610.00904", "title": "Screw Attention: Rigid-Body Algebra Inside a Transformer", "titleZh": "螺旋注意力：将刚体代数嵌入Transformer", "date": "2026-10-01", "year": 2026, "url": "https://arxiv.org/abs/2610.00904", "paperUrl": "https://arxiv.org/abs/2610.00904", "projectUrl": null, "codeUrl": null, "category": "机器人几何与控制", "tags": [ "螺旋理论", "等变网络", "结构化注意力", "仿真操作" ], "directions": [ "运动控制" ], "robotFilters": [ "Franka（型号未注明）" ], "tier": "recent", "summary": "将刚体间变换直接用于注意力消息传递，实现逐连杆坐标系变更鲁棒性。", "abstractZh": "Screw Attention以机器人连杆和场景物体为token，将相对位姿、关节螺旋和惯量作为已知几何关系。原文重点检验结构化表示，使用仿真器物体位姿而非视觉或语言输入，尚未完成实机部署。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "仿真测试姿态调整、可达接近方向、插接和LIBERO-Spatial十任务。LIBERO直接读连杆/物体位姿并给源—目标关系，不含图像及语言理解；2种子各500回合，对照含匹配参数Transformer、图网络和大型MLP。", "robots": [ "固定基座串联机械臂（仿真）", "Franka机械臂/LIBERO（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现需准确关节螺旋、相对变换与惯量，按任务选择力矩或末端twist读出。应严格保留无视觉/语言的条件、逐参考系变更协议和容量匹配；不能与标准图像输入VLA榜单直接比较成功率。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "每个token具有标量通道与六维运动/力螺旋通道；消息经真实相对变换传入接收坐标系，注意力只看不变量，学习矩阵仅混合通道。门控修正softmax归一化，使单层包含刚体速度递推；可独立输出动作或作为解析控制器的门控残差。", "whyUseful": "复现需准确关节螺旋、相对变换与惯量，按任务选择力矩或末端twist读出。应严格保留无视觉/语言的条件、逐参考系变更协议和容量匹配；不能与标准图像输入VLA榜单直接比较成功率。", "limitations": "所有结果来自仿真，未检验真实位姿估计、旋转误差、连杆长度或惯量误差。逐刚体单关节假设不覆盖浮动基座。少参数不等于快：CPU单次0.66毫秒，MLP为0.05毫秒；等变性也未单独降低示范需求。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00904v1", "note": "§IV-A，PDF页7–8：明确直接从仿真读取位姿，无相机或语言输入。" }, { "url": "https://arxiv.org/pdf/2610.00904v1", "note": "图6，PDF页10：97.3%来自2种子、每种子500回合；CPU推理并非最快。" }, { "url": "https://arxiv.org/pdf/2610.00904v1", "note": "§VI，PDF页12：明确仅仿真，浮动基座和旋转位姿误差未覆盖。" }, { "url": "https://arxiv.org/pdf/2610.00904v1", "note": "PDF页10（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823224+00:00", "original": { "id": "arxiv-2610.00904", "originalSourceUrl": "https://arxiv.org/pdf/2610.00904v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00904v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00904", "sha256": "80b8b57838ab9dafadeab2ec237fec3c9fc9597d0cd83afe080ee1af0b4c3493", "bytes": 1609601, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:14.760453+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00904", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00904", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00904", "pages": 17, "sourceIdentityStatus": "title_match", "sourceTitle": "Screw Attention: Rigid-Body Algebra Inside a Transformer", "sourceVersion": "2610.00904v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.574212+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00904" }, "originalAnalysis": { "id": "arxiv-2610.00904", "sectionsRead": [ "§III方法，PDF页3–7", "§IV-A–C，PDF页7–8", "§IV-F–H及图6–8，PDF页10–11", "§V–VI，PDF页11–12" ], "analyzedAt": "2026-10-04T13:58:28.823224+00:00", "analysisStatus": "full_text_sections", "methodsZh": "每个token具有标量通道与六维运动/力螺旋通道；消息经真实相对变换传入接收坐标系，注意力只看不变量，学习矩阵仅混合通道。门控修正softmax归一化，使单层包含刚体速度递推；可独立输出动作或作为解析控制器的门控残差。", "experimentsZh": "仿真测试姿态调整、可达接近方向、插接和LIBERO-Spatial十任务。LIBERO直接读连杆/物体位姿并给源—目标关系，不含图像及语言理解；2种子各500回合，对照含匹配参数Transformer、图网络和大型MLP。", "resultsZh": "16162参数模型在LIBERO达97.3%，大型MLP为95.2%；更换各token参考系后为97.7%，非等变对照降至约0–2.7%。插接残差较解析控制提高17.3点；10毫米位置噪声内表现稳定。", "limitationsZh": "所有结果来自仿真，未检验真实位姿估计、旋转误差、连杆长度或惯量误差。逐刚体单关节假设不覆盖浮动基座。少参数不等于快：CPU单次0.66毫秒，MLP为0.05毫秒；等变性也未单独降低示范需求。", "reproductionZh": "复现需准确关节螺旋、相对变换与惯量，按任务选择力矩或末端twist读出。应严格保留无视觉/语言的条件、逐参考系变更协议和容量匹配；不能与标准图像输入VLA榜单直接比较成功率。", "sourceUrl": "https://arxiv.org/pdf/2610.00904v1", "sourceVersion": "2610.00904v1", "originalSha256": "80b8b57838ab9dafadeab2ec237fec3c9fc9597d0cd83afe080ee1af0b4c3493", "experimentType": "sim", "robots": [ "固定基座串联机械臂（仿真）", "Franka机械臂/LIBERO（仿真）" ], "evidenceNotes": [ { "section": "§IV-A，PDF页7–8", "note": "明确直接从仿真读取位姿，无相机或语言输入。" }, { "section": "图6，PDF页10", "note": "97.3%来自2种子、每种子500回合；CPU推理并非最快。" }, { "section": "§VI，PDF页12", "note": "明确仅仿真，浮动基座和旋转位姿误差未覆盖。" }, { "section": "PDF页10（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "几何归纳偏置", "模型与学习结合" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Learned manipulation policies rediscover from data the spatial relations that rigid-body mechanics supplies in closed form. This costs data, and it leaves the policies fragile to geometric changes in the scene. We present Screw Attention, a transformer layer in which the relation between two bodies is a spatial transform rather than a graph edge. Every token is a body with a pose. Each pair of tokens carries the relative pose and, for robot joints, the joint screw. Messages are transported along this relation into the receiver's frame, while the attention scores see only frame-invariant quantities. By construction, the messages are equivariant to an independent change of frame at every token, and a single layer can express the velocity recursion of rigid-body mechanics. On simulated manipulation tasks, Screw Attention matches or outperforms controls of the same size, including graph, transformer and flat networks on LIBERO-Spatial. With 16,162 parameters it reaches 97.3% on LIBERO-Spatial from object poses (without images or language), above a flat network with 27x more parameters. Under a change of per-link frame convention its success is unchanged, while every other learned network falls below 3%. Placed on an analytic controller as a gated residual, it raises insertion success by 17.3 points. It is unaffected by pose noise up to 10,mm and by joint offsets within the factory calibration of a Franka arm. These results suggest a criterion: geometry is decisive when the task requires relations between frames that no other part of the system supplies. Code and trained policies will be released." }, { "id": "arxiv-2610.00913", "title": "eRLT: Efficient VLA Reinforcement Learning via Action-Relevant Token Routing", "titleZh": "eRLT：通过动作相关词元路由实现高效视觉—语言—动作模型强化学习", "date": "2026-10-01", "year": 2026, "url": "https://arxiv.org/abs/2610.00913", "paperUrl": "https://arxiv.org/abs/2610.00913", "projectUrl": null, "codeUrl": null, "category": "视觉语言动作模型", "tags": [ "在线强化学习", "词元路由", "精密插接", "冻结VLA" ], "directions": [ "视觉语言动作", "强化学习" ], "robotFilters": [ "RB-Y1" ], "tier": "recent", "summary": "跨词元与层提取动作相关表示，加快冻结VLA的在线适配。", "abstractZh": "eRLT不改动基础VLA，而为外接演员—评论家学习任务相关状态。示范动作预测初始化路由，在线评论家反馈再调整信息选择，在仿真和有人干预的真实插接中改善学习曲线。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "both", "experimentNote": "仿真覆盖4个LIBERO、3个RoboTwin任务并固定同设置内的算法与交互预算。实机使用RB-Y1左臂和双腕相机，USB及排线分别先微调99和800条示范，再采80和90条在线轨迹；失败会触发人工辅助。", "robots": [ "RB-Y1左臂（实机）", "LIBERO与RoboTwin任务机器人（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现须使用相同SFT检查点、回放/奖励/预算并保留原始观测以重算token。AUC用未平滑点梯形积分；应分别记自主与辅助轨迹，eRLT USB为66+14，排线70+20，避免把辅助成功计为自主性能。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "向冻结VLM额外追加可学习路由token，在多层读取其隐藏状态，再用任务共享层权重压缩为RL token。先用专家动作块预测监督初始化，移除预测头后通过评论家TD误差周期更新路由；演员梯度断开，基础VLA及其动作生成保持冻结。", "whyUseful": "复现须使用相同SFT检查点、回放/奖励/预算并保留原始观测以重算token。AUC用未平滑点梯形积分；应分别记自主与辅助轨迹，eRLT USB为66+14，排线70+20，避免把辅助成功计为自主性能。", "limitations": "优势主要是样本效率，不能把相对AUC提升当成功率百分点。实机仅两任务、单平台，辅助次数随失败变化，运行间方差和干预敏感性不足；额外VLM前向与周期重算增加显存和计算。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00913v1", "note": "§4，PDF页5–6：路由前向与原VLA动作生成分离；仅评论家更新路由。" }, { "url": "https://arxiv.org/pdf/2610.00913v1", "note": "表1、4，PDF页8–9：仿真均值与实机AUC、末10回合成功率为不同指标。" }, { "url": "https://arxiv.org/pdf/2610.00913v1", "note": "附录D.1表9，PDF页19–20：不同方法人工辅助数量不同；自主成功单独计数。" }, { "url": "https://arxiv.org/pdf/2610.00913v1", "note": "PDF页9（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823222+00:00", "original": { "id": "arxiv-2610.00913", "originalSourceUrl": "https://arxiv.org/pdf/2610.00913v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00913v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00913", "sha256": "99bd6e54b9370ad65d298e20902c8736c506ad9c24557da4726871d41e37d496", "bytes": 8904353, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:03.116411+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00913", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00913", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00913", "pages": 21, "sourceIdentityStatus": "title_match", "sourceTitle": "eRLT: Efficient VLA Reinforcement Learning via Action-Relevant Token Routing", "sourceVersion": "2610.00913v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.571670+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00913" }, "originalAnalysis": { "id": "arxiv-2610.00913", "sectionsRead": [ "§3.2–4，PDF页4–6", "§5及表1–4，PDF页7–9", "§6限制，PDF页10", "附录C.6、D.1及表9，PDF页19–20" ], "analyzedAt": "2026-10-04T13:58:28.823222+00:00", "analysisStatus": "full_text_sections", "methodsZh": "向冻结VLM额外追加可学习路由token，在多层读取其隐藏状态，再用任务共享层权重压缩为RL token。先用专家动作块预测监督初始化，移除预测头后通过评论家TD误差周期更新路由；演员梯度断开，基础VLA及其动作生成保持冻结。", "experimentsZh": "仿真覆盖4个LIBERO、3个RoboTwin任务并固定同设置内的算法与交互预算。实机使用RB-Y1左臂和双腕相机，USB及排线分别先微调99和800条示范，再采80和90条在线轨迹；失败会触发人工辅助。", "resultsZh": "七任务平均归一化AUC为0.626，对DSRL表征0.585、RLT表征0.506。USB AUC为0.562对0.269，末10次成功率90%对50%；排线AUC0.710对0.484，两者末窗均100%。", "limitationsZh": "优势主要是样本效率，不能把相对AUC提升当成功率百分点。实机仅两任务、单平台，辅助次数随失败变化，运行间方差和干预敏感性不足；额外VLM前向与周期重算增加显存和计算。", "reproductionZh": "复现须使用相同SFT检查点、回放/奖励/预算并保留原始观测以重算token。AUC用未平滑点梯形积分；应分别记自主与辅助轨迹，eRLT USB为66+14，排线70+20，避免把辅助成功计为自主性能。", "sourceUrl": "https://arxiv.org/pdf/2610.00913v1", "sourceVersion": "2610.00913v1", "originalSha256": "99bd6e54b9370ad65d298e20902c8736c506ad9c24557da4726871d41e37d496", "experimentType": "both", "robots": [ "RB-Y1左臂（实机）", "LIBERO与RoboTwin任务机器人（仿真）" ], "evidenceNotes": [ { "section": "§4，PDF页5–6", "note": "路由前向与原VLA动作生成分离；仅评论家更新路由。" }, { "section": "表1、4，PDF页8–9", "note": "仿真均值与实机AUC、末10回合成功率为不同指标。" }, { "section": "附录D.1表9，PDF页19–20", "note": "不同方法人工辅助数量不同；自主成功单独计数。" }, { "section": "PDF页9（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "VLA高效适配", "动作相关表征" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Vision-Language-Action (VLA) models provide strong behavioral priors for robotic manipulation, yet efficiently adapting them to downstream tasks remains challenging. Recent work addresses this challenge by adapting frozen VLAs through online reinforcement learning (RL), whose sample efficiency depends on the quality of the state representation used by the actor and critic. Existing methods construct such representations either with VLA-independent visual encoders or through fixed compression of internal VLA representations. Neither design explicitly extracts the task-specific action-relevant VLA features most useful for downstream action refinement and action-value estimation, therefore limiting sample efficiency. To address this limitation, we introduce eRLT, which constructs an effective state representation by routing task-specific action-relevant information across both tokens and layers of the frozen VLA. Specifically, learned routing tokens dynamically aggregate visual-language features at multiple depths, while a lightweight layer router combines these summaries into a fixed-dimensional RL token. The routing module is initialized using expert demonstrations to capture features predictive of expert actions and then refined using critic feedback from online interactions for action-value estimation. Across seven LIBERO and RoboTwin tasks, eRLT improves mean normalized learning-curve AUC by up to 23.7% over representative baselines. Real-robot experiments on USB connector insertion and motherboard ribbon-cable insertion further show AUC improvements of 108.9% and 46.7%, respectively, over the strongest baseline." }, { "id": "arxiv-2610.00921", "title": "In CEM, a World Model Is Also a Proposal Mechanism", "titleZh": "在交叉熵方法中，世界模型也是一种候选提案机制", "date": "2026-10-01", "year": 2026, "url": "https://arxiv.org/abs/2610.00921", "paperUrl": "https://arxiv.org/abs/2610.00921", "projectUrl": null, "codeUrl": null, "category": "世界模型与规划", "tags": [ "CEM", "世界模型评测", "模型预测控制", "因果干预" ], "directions": [ "世界模型" ], "robotFilters": [], "tier": "recent", "summary": "交叉重评分与单次精英集合替换揭示：候选分布收缩不代表模型排序更准确。", "abstractZh": "论文把世界模型在CEM中的两种作用拆开：为当前序列评分，以及通过精英更新改变下一轮搜索范围。作者在可重置仿真中执行所有候选建立参考，用交叉重评分和一次受控更新检验错误如何传递到后续搜索。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "DeepMind Control Suite的Walker walk与Cheetah run各训练6个独立Dreamer，共12个任务—种子单位。每单位4个状态，CEM为4轮、256候选、32精英、15步；完整交叉审计执行196608个序列，另留6单位检验预先选定干预。", "robots": [ "DeepMind Control Suite Walker（仿真）", "DeepMind Control Suite Cheetah（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现关键是精确保存/恢复仿真状态、候选ID、共同高斯随机样本及独立训练种子；统计单位是任务—种子而非数十万候选。应同时报告距离组成、精英恢复和真实成本，不能只复制一个聚合指标。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "保存各模型生成的候选池，再由四种固定代理模型交叉重评分；所有动作从同一仿真状态执行，获得真实成本和参考精英集。比较精英重合、成对排序，以及提案均值和宽度的不同归一化距离；仅替换首轮精英集并固定后续模型和随机样本，检验更新后果。", "whyUseful": "复现关键是精确保存/恢复仿真状态、候选ID、共同高斯随机样本及独立训练种子；统计单位是任务—种子而非数十万候选。应同时报告距离组成、精英恢复和真实成本，不能只复制一个聚合指标。", "limitations": "变化受距离定义和搜索宽度影响，不能把收缩误解为排序提高。只覆盖两任务、少量状态及单一规划配置；首轮干预使用部署时不可得的环境真值，且有限时域序列成本不等于闭环回报。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00921v1", "note": "§4–5，PDF页5–7：四模型交叉评分、12独立单位及196608候选；Identity宽度1280，其余256，非严格容量匹配。" }, { "url": "https://arxiv.org/pdf/2610.00921v1", "note": "§6.1–6.3，PDF页8–9：排序不改善；单轮替换最终成本在12单位均下降。" }, { "url": "https://arxiv.org/pdf/2610.00921v1", "note": "§7，PDF页10：明确定位为重置环境中的oracle诊断，不证明闭环收益。" }, { "url": "https://arxiv.org/pdf/2610.00921v1", "note": "PDF页8（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823218+00:00", "original": { "id": "arxiv-2610.00921", "originalSourceUrl": "https://arxiv.org/pdf/2610.00921v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00921v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00921", "sha256": "846471c39c478d54b094e6b616e1efbde96d0a596b0b881198676fda2002d677", "bytes": 572890, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:02.888294+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00921", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00921", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00921", "pages": 20, "sourceIdentityStatus": "title_match", "sourceTitle": "In CEM, a World Model Is Also a Proposal Mechanism", "sourceVersion": "2610.00921v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.564728+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00921" }, "originalAnalysis": { "id": "arxiv-2610.00921", "sectionsRead": [ "§3–5，PDF页3–7", "§6结果，PDF页8–9", "§7讨论与限制，PDF页10" ], "analyzedAt": "2026-10-04T13:58:28.823218+00:00", "analysisStatus": "full_text_sections", "methodsZh": "保存各模型生成的候选池，再由四种固定代理模型交叉重评分；所有动作从同一仿真状态执行，获得真实成本和参考精英集。比较精英重合、成对排序，以及提案均值和宽度的不同归一化距离；仅替换首轮精英集并固定后续模型和随机样本，检验更新后果。", "experimentsZh": "DeepMind Control Suite的Walker walk与Cheetah run各训练6个独立Dreamer，共12个任务—种子单位。每单位4个状态，CEM为4轮、256候选、32精英、15步；完整交叉审计执行196608个序列，另留6单位检验预先选定干预。", "resultsZh": "全部单位的预定提案距离都下降，但Walker排序约0.503近随机，Cheetah从0.535降至0.523。环境精英替换在12单位均降低最终所选序列成本；合并归一化改善为Walker −0.329、Cheetah −0.766。", "limitationsZh": "变化受距离定义和搜索宽度影响，不能把收缩误解为排序提高。只覆盖两任务、少量状态及单一规划配置；首轮干预使用部署时不可得的环境真值，且有限时域序列成本不等于闭环回报。", "reproductionZh": "复现关键是精确保存/恢复仿真状态、候选ID、共同高斯随机样本及独立训练种子；统计单位是任务—种子而非数十万候选。应同时报告距离组成、精英恢复和真实成本，不能只复制一个聚合指标。", "sourceUrl": "https://arxiv.org/pdf/2610.00921v1", "sourceVersion": "2610.00921v1", "originalSha256": "846471c39c478d54b094e6b616e1efbde96d0a596b0b881198676fda2002d677", "experimentType": "sim", "robots": [ "DeepMind Control Suite Walker（仿真）", "DeepMind Control Suite Cheetah（仿真）" ], "evidenceNotes": [ { "section": "§4–5，PDF页5–7", "note": "四模型交叉评分、12独立单位及196608候选；Identity宽度1280，其余256，非严格容量匹配。" }, { "section": "§6.1–6.3，PDF页8–9", "note": "排序不改善；单轮替换最终成本在12单位均下降。" }, { "section": "§7，PDF页10", "note": "明确定位为重置环境中的oracle诊断，不证明闭环收益。" }, { "section": "PDF页8（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "模型与规划器耦合诊断", "候选分布偏差" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "The cross-entropy method (CEM) uses world-model scores to select action sequences and fit the distribution sampled in its next iteration. A scoring error can therefore change both the present decision and the candidates considered later. We evaluate these two roles separately. Four types of predictive model generate CEM traces, and every model rescores every saved candidate pool. Executing the same candidates in the environment provides a reference elite set and proposal update. Across twelve independently trained task-seed units on Walker and Cheetah, the pre-specified proposal distance falls from the first to the final CEM iteration in every unit. Proposal widths contract and fitted means separate relative to the remaining search width. Pairwise ranking agreement stays near chance on Walker and declines on Cheetah; elite-set agreement does not improve. This comparison shows greater variation between scorers than between pool sources on Cheetah; Walker has variation in both and in their pairings. We use the original six units to select Random nonlinear for a one-update intervention, without inspecting intervention outcomes. Replacing its first model-ranked update with an environment-ranked update lowers final realised selected-sequence cost in those six units and in six further units held out from the selection." }, { "id": "arxiv-2610.00981", "title": "NarrativeFlow: Flow-Based Vision-Language-Action Model Using Robot Velocity Fields", "titleZh": "NarrativeFlow：利用机器人速度场的基于流的视觉—语言—动作模型", "date": "2026-10-01", "year": 2026, "url": "https://arxiv.org/abs/2610.00981", "paperUrl": "https://arxiv.org/abs/2610.00981", "projectUrl": null, "codeUrl": null, "category": "视觉语言动作模型", "tags": [ "VLA", "流匹配", "移动操作", "语言条件" ], "directions": [ "视觉语言动作" ], "robotFilters": [ "Toyota HSR", "Google Robot", "WidowX" ], "tier": "recent", "summary": "以连续二维速度场表示机器人运动，并用场景变化叙述监督提升语言条件操作。", "abstractZh": "NarrativeFlow将图像与指令映射为连续机器人运动流，训练时借助起始与目标图像的变化描述学习任务相关表示。生成流再交给下游动作策略；原文既评测离线轨迹预测，也在丰田HSR上执行移动操作。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "real", "experimentNote": "离线使用Fractal和Bridge V2，分别留出2000个测试片段，以ADE、FDE、LTDR评价末端区域。实机为11自由度丰田HSR，主要三任务每方法各20次、每任务约30条遥操作示范；附录扩展到13任务、每方法260次。", "robots": [ "Toyota HSR（实机）", "Google Robot（Fractal数据来源）", "WidowX（Bridge V2数据来源）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现须准备CoTracker3轨迹、经300张末端标注微调的Robot-SAM及离线叙述。原文给出约4.87亿可训练参数、RTX5090约6小时训练及62毫秒推理；实体平台迁移还需另训流条件策略。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "视觉与语言编码后使用两个专用token，分别驱动Flow-as-Flow生成器和Narrative Delta辅助损失。后者对齐大模型依据起始、目标图像生成的多视角变化叙述；推理不需要目标图。DiT预测连续二维速度场，经积分得到运动点轨迹，再条件化操作策略。", "whyUseful": "复现须准备CoTracker3轨迹、经300张末端标注微调的Robot-SAM及离线叙述。原文给出约4.87亿可训练参数、RTX5090约6小时训练及62毫秒推理；实体平台迁移还需另训流条件策略。", "limitations": "假设固定相机且初始末端可见；二维流难表达沿视线方向的深度运动。杯堆叠仅20%且真值流上限同为20%，说明下游策略也是瓶颈。物理一致性主要依赖经验结果，不等于三维动力学保证。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00981v1", "note": "§3.1，PDF页5：明确假设固定相机、初始末端可见，并将主要研究聚焦于流生成。" }, { "url": "https://arxiv.org/pdf/2610.00981v1", "note": "表2，PDF页9；表4，PDF页13：离线ADE及HSR三任务成功率来自原文定量表。" }, { "url": "https://arxiv.org/pdf/2610.00981v1", "note": "附录D.2表D，PDF页23：13任务260次/方法，均值58%与45%；不将数据集采集机器人误当实机评测平台。" }, { "url": "https://arxiv.org/pdf/2610.00981v1", "note": "PDF页13（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823202+00:00", "original": { "id": "arxiv-2610.00981", "originalSourceUrl": "https://arxiv.org/pdf/2610.00981v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00981v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00981", "sha256": "9e15201052d0fc948b2497c7643afe2406831f3b825b8114327c5c788427b6bb", "bytes": 5040404, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:03.021296+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00981", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00981", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00981", "pages": 24, "sourceIdentityStatus": "title_match", "sourceTitle": "NarrativeFlow: Flow-Based Vision-Language-Action Model Using Robot Velocity Fields", "sourceVersion": "2610.00981v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.563264+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00981" }, "originalAnalysis": { "id": "arxiv-2610.00981", "sectionsRead": [ "§3方法，PDF页4–7", "§4.1–4.2，PDF页8–9", "§5及表3–4，PDF页12–14", "附录D.2及表D，PDF页23" ], "analyzedAt": "2026-10-04T13:58:28.823202+00:00", "analysisStatus": "full_text_sections", "methodsZh": "视觉与语言编码后使用两个专用token，分别驱动Flow-as-Flow生成器和Narrative Delta辅助损失。后者对齐大模型依据起始、目标图像生成的多视角变化叙述；推理不需要目标图。DiT预测连续二维速度场，经积分得到运动点轨迹，再条件化操作策略。", "experimentsZh": "离线使用Fractal和Bridge V2，分别留出2000个测试片段，以ADE、FDE、LTDR评价末端区域。实机为11自由度丰田HSR，主要三任务每方法各20次、每任务约30条遥操作示范；附录扩展到13任务、每方法260次。", "resultsZh": "Fractal/Bridge V2的ADE为21.68/30.59，优于两个基线。三项实机任务平均成功率55%，Im2Flow2Act为42%；13任务扩展均值58%对45%。分离token和叙述监督的消融均支持设计贡献。", "limitationsZh": "假设固定相机且初始末端可见；二维流难表达沿视线方向的深度运动。杯堆叠仅20%且真值流上限同为20%，说明下游策略也是瓶颈。物理一致性主要依赖经验结果，不等于三维动力学保证。", "reproductionZh": "复现须准备CoTracker3轨迹、经300张末端标注微调的Robot-SAM及离线叙述。原文给出约4.87亿可训练参数、RTX5090约6小时训练及62毫秒推理；实体平台迁移还需另训流条件策略。", "sourceUrl": "https://arxiv.org/pdf/2610.00981v1", "sourceVersion": "2610.00981v1", "originalSha256": "9e15201052d0fc948b2497c7643afe2406831f3b825b8114327c5c788427b6bb", "experimentType": "real", "robots": [ "Toyota HSR（实机）", "Google Robot（Fractal数据来源）", "WidowX（Bridge V2数据来源）" ], "evidenceNotes": [ { "section": "§3.1，PDF页5", "note": "明确假设固定相机、初始末端可见，并将主要研究聚焦于流生成。" }, { "section": "表2，PDF页9；表4，PDF页13", "note": "离线ADE及HSR三任务成功率来自原文定量表。" }, { "section": "附录D.2表D，PDF页23", "note": "13任务260次/方法，均值58%与45%；不将数据集采集机器人误当实机评测平台。" }, { "section": "PDF页13（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "跨具身动作表示", "语言条件轨迹生成" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "We focus on language-conditioned flow-based manipulation, where robot flows (robot velocity fields) serve as embodiment-agnostic, motion-centric representations for leveraging data collected from multiple robot platforms. This task is crucial because language-conditioned manipulation is essential for practical robotic systems, yet scaling robot foundation models remains limited by the labor-intensive collection of embodiment-specific data. Existing methods either coarsely approximate robot flows with sparse keypoint displacements, or cannot handle language-conditioned manipulation. To address this limitation, we propose NarrativeFlow, which models robot flows as continuous velocity fields using a flow-matching formulation conditioned on language. Accordingly, NarrativeFlow generates robot flows that are physically consistent with real-world manipulation. To validate NarrativeFlow, we have conducted experiments on standard datasets for language-conditioned manipulation. The experimental results show that NarrativeFlow outperforms representative baseline methods on standard evaluation metrics. Furthermore, through real-world experiments, we show that NarrativeFlow achieves higher success rates than baseline methods across multiple manipulation tasks. The project page is available at https://shota0520.github.io/NarrativeFlow-project-page/" }, { "id": "arxiv-2610.00982", "title": "Divide-and-Remember: Recursive Action-Relevant Memory for Long-Horizon VLA Policies", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.00982v1", "titleZh": "Divide-and-Remember：面向长时程视觉—语言—动作策略的递归动作相关记忆", "abstractZh": "保留预训练视觉token本身，以共享轻量选择器递归从2K候选选K，组合长历史；用直通梯度和Gumbel探索训练选择，动作流匹配损失驱动记忆，无额外未来状态预测目标。递归结构使同一历史前缀在离线训练和在线执行中保持一致。\nRoboMME含16任务，固定64token预算、每任务50回合、三种子和末三检查点。真机使用Franka Panda加Robotiq2F-140，四任务各100示范、各10测试；前相机生成记忆，腕相机只提供当前观测。", "summary": "保留预训练视觉token本身，以共享轻量选择器递归从2K候选选K，组合长历史；用直通梯度和Gumbel探索训练选择，动作流匹配损失驱动记忆，无额外未来状态预测目标。递归结构使同一历史前缀在离线训练和在线执行中保…", "experimentType": "both", "robots": [ "Franka Emika Panda" ], "tags": [ "视觉语言动作模型", "长时程操作", "记忆机制", "模仿学习" ], "limitations": "真机比较预算不同：D&R1024，FrameSamp2048，不能称真机也固定64token。InsertPeg所有方法最高仅3.3%，记忆不解决接触精度；较大记忆有收敛与冗余问题，四任务小样本也不代表泛化可靠性。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.00982v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.00982v1", "note": "§6.1：仿真同64token；800回合，三种子×三检查点。" }, { "url": "https://arxiv.org/html/2610.00982v1", "note": "Appendix B.4：Panda及Robotiq夹爪；真机1024对2048token，不等预算。" }, { "url": "https://arxiv.org/html/2610.00982v1", "note": "Table 2：实机35/40，不是仿真38.6%。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "视觉语言动作模型 / 长时程操作", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "视觉语言动作", "操作与抓取", "模仿学习" ], "robotFilters": [ "Franka Panda" ], "original": { "id": "arxiv-2610.00982", "originalSourceUrl": "https://arxiv.org/pdf/2610.00982v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00982v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00982v1", "sha256": "31070beb3c4ab4ff1fa166c04b920b5c4816ef220b88729147b9405e15be3346", "bytes": 8606466, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:11.356301+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00982v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00982v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00982v1", "pages": 27, "sourceIdentityStatus": "title_match", "sourceTitle": "Divide-and-Remember: Recursive Action-Relevant Memory for Long-Horizon VLA Policies", "sourceVersion": "2610.00982v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.591199+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00982v1" }, "originalAnalysis": { "id": "arxiv-2610.00982", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090121+00:00", "sourceUrl": "https://arxiv.org/html/2610.00982v1", "sectionsRead": [ "§4–5", "§6.1–6.3", "Tables 1–2", "§7", "Appendix B.4" ], "methodsZh": "保留预训练视觉token本身，以共享轻量选择器递归从2K候选选K，组合长历史；用直通梯度和Gumbel探索训练选择，动作流匹配损失驱动记忆，无额外未来状态预测目标。递归结构使同一历史前缀在离线训练和在线执行中保持一致。", "experimentsZh": "RoboMME含16任务，固定64token预算、每任务50回合、三种子和末三检查点。真机使用Franka Panda加Robotiq2F-140，四任务各100示范、各10测试；前相机生成记忆，腕相机只提供当前观测。", "resultsZh": "作者报告四套件平均成功38.6%，HAMLET32.2%、无记忆17.9%；实体测试35/40，FrameSamp22/40、无记忆3/40。帧式事实收益较大，但时间与动态状态类未全面超过潜在记忆。", "limitationsZh": "真机比较预算不同：D&R1024，FrameSamp2048，不能称真机也固定64token。InsertPeg所有方法最高仅3.3%，记忆不解决接触精度；较大记忆有收敛与冗余问题，四任务小样本也不代表泛化可靠性。", "reproductionZh": "复现需固定候选池采样、冻结SigLIP与共同策略入口，并区分套件均值和任务均值；实机动作/相机时间同步及停止判定影响得分。文中指向代码/检查点，但本轮未实训或复现性能。", "experimentType": "both", "robots": [ "Franka Emika Panda" ], "evidenceNotes": [ { "section/page": "§6.1", "note": "仿真同64token；800回合，三种子×三检查点。" }, { "section/page": "Appendix B.4", "note": "Panda及Robotiq夹爪；真机1024对2048token，不等预算。" }, { "section/page": "Table 2", "note": "实机35/40，不是仿真38.6%。" } ], "corrections": { "robots": [ "Franka Emika Panda" ], "budgetNote": "仅仿真为等64token预算；真机D&R1024、FrameSamp2048。" } }, "experimentNote": "RoboMME含16任务，固定64token预算、每任务50回合、三种子和末三检查点。真机使用Franka Panda加Robotiq2F-140，四任务各100示范、各10测试；前相机生成记忆，腕相机只提供当前观测。", "contribution": "保留预训练视觉token本身，以共享轻量选择器递归从2K候选选K，组合长历史；用直通梯度和Gumbel探索训练选择，动作流匹配损失驱动记忆，无额外未来状态预测目标。递归结构使同一历史前缀在离线训练和在线执行中保持一致。", "whyUseful": "复现需固定候选池采样、冻结SigLIP与共同策略入口，并区分套件均值和任务均值；实机动作/相机时间同步及停止判定影响得分。文中指向代码/检查点，但本轮未实训或复现性能。", "analysisVerifiedAt": "2026-10-04T13:50:35.090121+00:00" }, { "id": "arxiv-2610.01004", "title": "The Effect of Gait Stability Based on Two Types of Impact Strategies for Two-Link Walking and Brachiating Robots", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01004v1", "titleZh": "两种冲击切换策略对双连杆行走与臂行机器人的步态稳定性影响", "abstractZh": "论文将双连杆被动行走和臂摆统一为连续双摆动力学、瞬时完全塑性碰撞、坐标重标记组成的步间映射。比较足/手碰到表面即触发的状态切换SBS与固定间隔触发的时间切换TBS；两者可拥有相同周期一步轨道，但扰动后的碰撞时刻不同，因此稳定性不必相同。\n固定一组无量纲形态参数，以步长0.001的伪弧长延拓追踪周期解，最多50,000点；用线性化冲击映射特征值判断局部稳定，对不稳定轨道再数值演化1000次碰撞，分析分岔和镜像步态的吸引域。", "summary": "论文将双连杆被动行走和臂摆统一为连续双摆动力学、瞬时完全塑性碰撞、坐标重标记组成的步间映射。比较足/手碰到表面即触发的状态切换SBS与固定间隔触发的时间切换TBS；两者可拥有相同周期一步轨道，但扰动后的碰撞时刻…", "experimentType": "sim", "robots": [ "双连杆行走/臂摆数学模型" ], "tags": [ "步态稳定性", "冲击动力学", "双连杆模型", "分岔分析" ], "limitations": "这是理想化被动模型数值研究，没有实体机器人或学习策略测试。完全塑性接触及允许上下表面切换是关键假设，跌落后的收敛不是现实安全恢复保证。作者未正式分析混沌，结论也不能推广到任意形态。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01004v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01004v1", "note": "2：被动双摆与瞬时塑性冲击模型。" }, { "url": "https://arxiv.org/html/2610.01004v1", "note": "3：0.001延拓步长，最多50,000点，1000次冲击分析。" }, { "url": "https://arxiv.org/html/2610.01004v1", "note": "3.2; 4：TBS稳定区间与临界值文字有歧义；混沌正式分析留待以后。" }, { "url": "https://arxiv.org/html/2610.01004v1", "note": "PDF p7, Figure 4：已渲染核对SBS倍周期与TBS闭合不变曲线图；正文稳定时间区间与临界值存在内部歧义。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "步态稳定性 / 冲击动力学", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "运动控制" ], "robotFilters": [], "original": { "id": "arxiv-2610.01004", "originalSourceUrl": "https://arxiv.org/pdf/2610.01004v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01004v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01004v1", "sha256": "f796a8d4eecf2999a452eb1f23e81be9854a498dd1fa58763c8c346d9df23ef5", "bytes": 9377286, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:04.963300+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01004v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01004v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01004v1", "pages": 10, "sourceIdentityStatus": "title_match", "sourceTitle": "The Effect of Gait Stability Based on Two Types of Impact Strategies for Two-Link Walking and Brachiating Robots", "sourceVersion": "2610.01004v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.583980+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01004v1" }, "originalAnalysis": { "id": "arxiv-2610.01004", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458236+00:00", "sourceUrl": "https://arxiv.org/html/2610.01004v1", "sectionsRead": [ "2.1–2.3", "3.1–3.4", "Figures 3–6", "4 Conclusion" ], "methodsZh": "论文将双连杆被动行走和臂摆统一为连续双摆动力学、瞬时完全塑性碰撞、坐标重标记组成的步间映射。比较足/手碰到表面即触发的状态切换SBS与固定间隔触发的时间切换TBS；两者可拥有相同周期一步轨道，但扰动后的碰撞时刻不同，因此稳定性不必相同。", "experimentsZh": "固定一组无量纲形态参数，以步长0.001的伪弧长延拓追踪周期解，最多50,000点；用线性化冲击映射特征值判断局部稳定，对不稳定轨道再数值演化1000次碰撞，分析分岔和镜像步态的吸引域。", "resultsZh": "作者在所研究参数下观察到SBS的折叠、倍周期及Neimark–Sacker分岔，TBS以后两类中的折叠和Neimark–Sacker为主，稳定区域更窄。部分不稳定行走轨迹跌落后转为稳定臂摆；TBS还可能收敛到反方向镜像家族。", "limitationsZh": "这是理想化被动模型数值研究，没有实体机器人或学习策略测试。完全塑性接触及允许上下表面切换是关键假设，跌落后的收敛不是现实安全恢复保证。作者未正式分析混沌，结论也不能推广到任意形态。", "reproductionZh": "论文提供TWOLINK_NODYCON代码地址；复现重点是无量纲参数、事件检测、冲击雅可比、延拓初始解及步长，而非训练网络。原文3.2的TBS稳定区间与所列临界值表述不完全一致，精确边界应核对图和代码。", "experimentType": "sim", "robots": [ "双连杆行走/臂摆数学模型" ], "evidenceNotes": [ { "section": "2", "note": "被动双摆与瞬时塑性冲击模型。" }, { "section": "3", "note": "0.001延拓步长，最多50,000点，1000次冲击分析。" }, { "section": "3.2; 4", "note": "TBS稳定区间与临界值文字有歧义；混沌正式分析留待以后。" }, { "section": "PDF p7, Figure 4", "note": "已渲染核对SBS倍周期与TBS闭合不变曲线图；正文稳定时间区间与临界值存在内部歧义。" } ], "corrections": [ "无实机硬件验证；步态跌落后恢复只成立于所设数值模型。" ], "sourceVersion": "2610.01004v1", "originalSha256": "f796a8d4eecf2999a452eb1f23e81be9854a498dd1fa58763c8c346d9df23ef5" }, "experimentNote": "固定一组无量纲形态参数，以步长0.001的伪弧长延拓追踪周期解，最多50,000点；用线性化冲击映射特征值判断局部稳定，对不稳定轨道再数值演化1000次碰撞，分析分岔和镜像步态的吸引域。", "contribution": "论文将双连杆被动行走和臂摆统一为连续双摆动力学、瞬时完全塑性碰撞、坐标重标记组成的步间映射。比较足/手碰到表面即触发的状态切换SBS与固定间隔触发的时间切换TBS；两者可拥有相同周期一步轨道，但扰动后的碰撞时刻不同，因此稳定性不必相同。", "whyUseful": "论文提供TWOLINK_NODYCON代码地址；复现重点是无量纲参数、事件检测、冲击雅可比、延拓初始解及步长，而非训练网络。原文3.2的TBS稳定区间与所列临界值表述不完全一致，精确边界应核对图和代码。", "analysisVerifiedAt": "2026-10-04T13:45:34.458236+00:00" }, { "id": "arxiv-2610.01019", "title": "FutureWorlds: Learning Robotic World Models from Alternative Futures", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01019v1", "titleZh": "FutureWorlds：从多种备选未来中学习机器人世界模型", "abstractZh": "用离散自回归视觉模型预测固定动作序列下的多个未来；多样化beam search产生候选，每个候选单独保留初始锚点与滚动历史。MemSPO按视频轨迹奖励计算组内优势，生成和概率打分使用同一历史，避免候选分叉后记忆错配。\nRT-1、BridgeV2和RoboCasa各128条固定留出轨迹，比较32帧预测的PSNR、SSIM、LPIPS及光流；固定解码与权重做记忆消融，并测试200次后训练更新。另有冻结RT-1策略在生成世界内关抽屉的单个定性例子。", "summary": "用离散自回归视觉模型预测固定动作序列下的多个未来；多样化beam search产生候选，每个候选单独保留初始锚点与滚动历史。MemSPO按视频轨迹奖励计算组内优势，生成和概率打分使用同一历史，避免候选分叉后记忆…", "experimentType": "data", "robots": [ "未核实" ], "tags": [ "世界模型", "视频预测", "强化学习", "记忆机制" ], "limitations": "主要是离线视频及学习模型内部反馈评估，没有实体机器人闭环任务成功证据；视频指标改善不能等同物理正确。beam筛选改变采样分布，作者明确不保证无偏同策略梯度；不同模型原生推理设置也不完全相同。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01019v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/pdf/2610.01019", "note": "3.4; 4.1–4.5：搜索偏差声明、固定评估、视觉与动态指标。" }, { "url": "https://arxiv.org/pdf/2610.01019", "note": "Appendix D; 5; Appendix G：生成世界单例不确立物理成功，资源与延迟范围。" }, { "url": "https://arxiv.org/pdf/2610.01019v1", "note": "3.4; 4.1–4.5：搜索偏差声明、固定评估、视觉与动态指标。" }, { "url": "https://arxiv.org/pdf/2610.01019v1", "note": "Appendix D; 5; Appendix G：生成世界单例不确立物理成功，资源与延迟范围。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "世界模型 / 视频预测", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "世界模型", "强化学习" ], "robotFilters": [], "original": { "id": "arxiv-2610.01019", "originalSourceUrl": "https://arxiv.org/pdf/2610.01019v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01019v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01019v1", "sha256": "c6d35ba6d0b05f199ecc50bcb4899190e9e9ee09c21237b1f7a96d148064bc87", "bytes": 33680170, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:00.390254+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01019v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01019v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01019v1", "pages": 32, "sourceIdentityStatus": "title_match", "sourceTitle": "FutureWorlds: Learning Robotic World Models from Alternative Futures", "sourceVersion": "2610.01019v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.574996+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01019v1" }, "originalAnalysis": { "id": "arxiv-2610.01019", "sourceUrl": "https://arxiv.org/pdf/2610.01019v1", "sectionsRead": [ "3.1–3.4", "4.1–4.5", "5", "Appendix D,F,G", "Tables 1–2" ], "methodsZh": "用离散自回归视觉模型预测固定动作序列下的多个未来；多样化beam search产生候选，每个候选单独保留初始锚点与滚动历史。MemSPO按视频轨迹奖励计算组内优势，生成和概率打分使用同一历史，避免候选分叉后记忆错配。", "experimentsZh": "RT-1、BridgeV2和RoboCasa各128条固定留出轨迹，比较32帧预测的PSNR、SSIM、LPIPS及光流；固定解码与权重做记忆消融，并测试200次后训练更新。另有冻结RT-1策略在生成世界内关抽屉的单个定性例子。", "resultsZh": "LPIPS相对各集最强基线下降14.78%、20.84%、9.12%；去掉历史使32帧MSE明显恶化，多数收益来自近期历史。含视觉编解码器约2.57亿参数、1.89GiB峰值显存，但32帧beam推理26.76秒，比iVideoGPT的4.14秒慢。", "limitationsZh": "主要是离线视频及学习模型内部反馈评估，没有实体机器人闭环任务成功证据；视频指标改善不能等同物理正确。beam筛选改变采样分布，作者明确不保证无偏同策略梯度；不同模型原生推理设置也不完全相同。", "reproductionZh": "先锁定128条轨迹、预测帧和缩放规则，分别复现SFT、普通beam后训练和多样搜索。必须记录候选记忆、奖励组件、CPU/GPU解码及延迟；不能把小参数量直接解释为实时机器人速度。", "experimentType": "data", "robots": [], "corrections": [ "从unknown改为data：真实机器人数据集不等于本研究执行了实机闭环评估。" ], "evidenceNotes": [ { "note": "搜索偏差声明、固定评估、视觉与动态指标。", "section": "3.4; 4.1–4.5" }, { "note": "生成世界单例不确立物理成功，资源与延迟范围。", "section": "Appendix D; 5; Appendix G" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2610.01019v1", "originalSha256": "c6d35ba6d0b05f199ecc50bcb4899190e9e9ee09c21237b1f7a96d148064bc87" }, "experimentNote": "RT-1、BridgeV2和RoboCasa各128条固定留出轨迹，比较32帧预测的PSNR、SSIM、LPIPS及光流；固定解码与权重做记忆消融，并测试200次后训练更新。另有冻结RT-1策略在生成世界内关抽屉的单个定性例子。", "contribution": "用离散自回归视觉模型预测固定动作序列下的多个未来；多样化beam search产生候选，每个候选单独保留初始锚点与滚动历史。MemSPO按视频轨迹奖励计算组内优势，生成和概率打分使用同一历史，避免候选分叉后记忆错配。", "whyUseful": "先锁定128条轨迹、预测帧和缩放规则，分别复现SFT、普通beam后训练和多样搜索。必须记录候选记忆、奖励组件、CPU/GPU解码及延迟；不能把小参数量直接解释为实时机器人速度。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2610.01072", "title": "quARtet Marker: A 3D-Printable Multi-Tag Fiducial for Robust Near-Frontal Pose Estimation", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01072v1", "titleZh": "quARtet Marker：用于稳健近正面位姿估计的可3D打印多标签基准标记", "abstractZh": "在35毫米方形底座上布置四枚倾斜AprilTag，用共享参数同时生成打印几何和PnP角点模型，缓解正视单平面标记的姿态歧义。比较对角、俯仰和反向俯仰三种布置，兼顾可见性及平行夹爪的接触平面。\nUR5e与固定相机分别扫描25个朝向和25个平移位置，每位姿十张连续图；另以腕载相机做每标记五次约117秒姿态保持，以及五次独立重新抓握摆落测试。补充仿真讨论倾斜角选择，静态帧属于技术重复。", "summary": "在35毫米方形底座上布置四枚倾斜AprilTag，用共享参数同时生成打印几何和PnP角点模型，缓解正视单平面标记的姿态歧义。比较对角、俯仰和反向俯仰三种布置，兼顾可见性及平行夹爪的接触平面。", "experimentType": "both", "robots": [ "Universal Robots UR5e" ], "tags": [ "视觉定位", "基准标记", "机器人抓取", "实验室自动化" ], "limitations": "误差相对机器人操作参考并经过常量配准，未建立绝对计量精度；连续十帧不是十次独立装配。斜视时部分构型劣于平面标签，抓握结论限所测夹爪、载荷及摆动，未评价检测运行时间。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01072v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/pdf/2610.01072v1", "note": "§4.1–4.2; Table 2：35毫米标签、UR5e、连续帧与独立五次闭环的区分。" }, { "url": "https://arxiv.org/pdf/2610.01072v1", "note": "Table 3; §5：对角型严重滑移及相对参考精度限制。" }, { "url": "https://arxiv.org/pdf/2610.01072v1", "note": "Supplement S1 pp32–34：18度倾角由理想单平面标签的像素扰动仿真辅助选定；不是完整多标签成像误差模型。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "视觉定位 / 基准标记", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "操作与抓取", "数据集与基准" ], "robotFilters": [ "Universal Robots UR5e" ], "original": { "id": "arxiv-2610.01072", "originalSourceUrl": "https://arxiv.org/pdf/2610.01072v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01072v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01072v1", "sha256": "333a51f18e859636881d33e45df987399e95f7b054b05edb1440eb55e9887c67", "bytes": 18873096, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:54.754576+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01072v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01072v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01072v1", "pages": 58, "sourceIdentityStatus": "title_match", "sourceTitle": "quARtet Marker: A 3D-Printable Multi-Tag Fiducial for Robust Near-Frontal Pose Estimation", "sourceVersion": "2610.01072v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.542382+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01072v1" }, "originalAnalysis": { "id": "arxiv-2610.01072", "sourceUrl": "https://arxiv.org/pdf/2610.01072v1", "sectionsRead": [ "§3", "§4.1–4.3", "Tables 2–3", "§5", "Supplement S1" ], "methodsZh": "在35毫米方形底座上布置四枚倾斜AprilTag，用共享参数同时生成打印几何和PnP角点模型，缓解正视单平面标记的姿态歧义。比较对角、俯仰和反向俯仰三种布置，兼顾可见性及平行夹爪的接触平面。", "experimentsZh": "UR5e与固定相机分别扫描25个朝向和25个平移位置，每位姿十张连续图；另以腕载相机做每标记五次约117秒姿态保持，以及五次独立重新抓握摆落测试。补充仿真讨论倾斜角选择，静态帧属于技术重复。", "resultsZh": "作者报告正视角误差由单标记2.18度降至0.24–0.47度，平移RMSE从1.50降到0.17–0.20毫米；闭环角误差由0.68降至0.08–0.11度。对角型抓握滑移约103.5毫米，俯仰型约1.7–2.0毫米，显示定位与抓握取舍。", "limitationsZh": "误差相对机器人操作参考并经过常量配准，未建立绝对计量精度；连续十帧不是十次独立装配。斜视时部分构型劣于平面标签，抓握结论限所测夹爪、载荷及摆动，未评价检测运行时间。", "reproductionZh": "需保存打印配置、角点坐标、相机/手眼标定及配准约定，统计以独立运行作单位。不得把对角型未完全掉落写成稳定抓持，或把亚毫米相对误差泛化为绝对精度；本次未打印测试。", "experimentType": "both", "robots": [ "Universal Robots UR5e" ], "evidenceNotes": [ { "section/page": "§4.1–4.2; Table 2", "note": "35毫米标签、UR5e、连续帧与独立五次闭环的区分。" }, { "section/page": "Table 3; §5", "note": "对角型严重滑移及相对参考精度限制。" }, { "section/page": "Supplement S1 pp32–34", "note": "18度倾角由理想单平面标签的像素扰动仿真辅助选定；不是完整多标签成像误差模型。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "UR5e与固定相机分别扫描25个朝向和25个平移位置，每位姿十张连续图；另以腕载相机做每标记五次约117秒姿态保持，以及五次独立重新抓握摆落测试。补充仿真讨论倾斜角选择，静态帧属于技术重复。", "contribution": "在35毫米方形底座上布置四枚倾斜AprilTag，用共享参数同时生成打印几何和PnP角点模型，缓解正视单平面标记的姿态歧义。比较对角、俯仰和反向俯仰三种布置，兼顾可见性及平行夹爪的接触平面。", "whyUseful": "需保存打印配置、角点坐标、相机/手眼标定及配准约定，统计以独立运行作单位。不得把对角型未完全掉落写成稳定抓持，或把亚毫米相对误差泛化为绝对精度；本次未打印测试。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2610.01083", "title": "WBAG: A Whole-Body and Attached-Geometry Safety Framework for Vision-Language-Action Manipulation", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01083v1", "titleZh": "WBAG：面向视觉—语言—动作操作的机器人全身及附着物几何安全框架", "abstractZh": "将机器人各运动连杆、障碍物和抓取后附着物建成可微Bernstein多项式SDF，以抓取模式切换受保护几何。对冻结π0.5的六维操作空间增量求最小改动CBF二次规划，保留原控制器和夹爪指令，无需重训策略。\n仅在SafeLIBERO仿真中测16任务、两种障碍布置、各50随机回合，共1600回合。各配置共享策略和初始条件，逐项比较单末端、全场景、全身及附着物模型；目标/支撑物身份由模拟器提供，隔离语义识别误差。", "summary": "将机器人各运动连杆、障碍物和抓取后附着物建成可微Bernstein多项式SDF，以抓取模式切换受保护几何。对冻结π0.5的六维操作空间增量求最小改动CBF二次规划，保留原控制器和夹爪指令，无需重训策略。", "experimentType": "sim", "robots": [ "未核实" ], "tags": [ "视觉语言动作模型", "安全控制", "碰撞避免", "控制障碍函数" ], "limitations": "安全指标是非任务物体位移超过1厘米的代理量，不能发现所有碰撞伤害。理论前向不变性依赖精确最近点和零松弛，而实现用有限点近似且允许松弛；夹爪开合运动也未纳入动作映射，因此不是无条件安全保证。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01083v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01083v1", "note": "§IV-E：作者明确区分零松弛定理和允许松弛的实际过滤器。" }, { "url": "https://arxiv.org/html/2610.01083v1", "note": "§V-A：1600仿真回合；模拟器给定任务身份；1厘米位移代理。" }, { "url": "https://arxiv.org/html/2610.01083v1", "note": "Table II：97.38%为Scene Safety，59.38%为Safe Success。" }, { "url": "https://arxiv.org/html/2610.01083v1", "note": "PDF p.7, Table II (visual inspection)：已核对97.38%安全、59.38%安全成功；Goal和Spatial的安全成功最佳不属WBAG。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "视觉语言动作模型 / 安全控制", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "视觉语言动作", "运动控制", "操作与抓取", "安全与评估" ], "robotFilters": [], "original": { "id": "arxiv-2610.01083", "originalSourceUrl": "https://arxiv.org/pdf/2610.01083v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01083v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01083v1", "sha256": "ffb75d2e97fbbb22bee133a75b27403e4f4d2951f6029bbeb3db904f4a7c1307", "bytes": 8440307, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:54.502348+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01083v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01083v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01083v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "WBAG: A Whole-Body and Attached-Geometry Safety Framework for Vision-Language-Action Manipulation", "sourceVersion": "2610.01083v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.527004+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01083v1" }, "originalAnalysis": { "id": "arxiv-2610.01083", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090120+00:00", "sourceUrl": "https://arxiv.org/html/2610.01083v1", "sectionsRead": [ "§II–IV", "§V", "Tables I–III", "Proposition 1" ], "methodsZh": "将机器人各运动连杆、障碍物和抓取后附着物建成可微Bernstein多项式SDF，以抓取模式切换受保护几何。对冻结π0.5的六维操作空间增量求最小改动CBF二次规划，保留原控制器和夹爪指令，无需重训策略。", "experimentsZh": "仅在SafeLIBERO仿真中测16任务、两种障碍布置、各50随机回合，共1600回合。各配置共享策略和初始条件，逐项比较单末端、全场景、全身及附着物模型；目标/支撑物身份由模拟器提供，隔离语义识别误差。", "resultsZh": "作者报告全场景安全97.38%、安全成功59.38%，AEGIS为70.87%/51.06%；但Goal、Spatial任务安全成功并非最佳。完整过滤器平均26.76毫秒，主要成本是接触点搜索。", "limitationsZh": "安全指标是非任务物体位移超过1厘米的代理量，不能发现所有碰撞伤害。理论前向不变性依赖精确最近点和零松弛，而实现用有限点近似且允许松弛；夹爪开合运动也未纳入动作映射，因此不是无条件安全保证。", "reproductionZh": "复现必须公开任务角色排除、对象身份元数据、SDF拟合、接触搜索、松弛惩罚及控制缩放，分别验证几何误差与真实控制执行。正文没有实体机器人部署证据，不据LIBERO默认模型补填实机型号。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "§IV-E", "note": "作者明确区分零松弛定理和允许松弛的实际过滤器。" }, { "section/page": "§V-A", "note": "1600仿真回合；模拟器给定任务身份；1厘米位移代理。" }, { "section/page": "Table II", "note": "97.38%为Scene Safety，59.38%为Safe Success。" }, { "section/page": "PDF p.7, Table II (visual inspection)", "note": "已核对97.38%安全、59.38%安全成功；Goal和Spatial的安全成功最佳不属WBAG。" } ], "corrections": { "safetyScope": "仅仿真位移代理指标；近似/松弛实现不具有无条件CBF安全保证。" } }, "experimentNote": "仅在SafeLIBERO仿真中测16任务、两种障碍布置、各50随机回合，共1600回合。各配置共享策略和初始条件，逐项比较单末端、全场景、全身及附着物模型；目标/支撑物身份由模拟器提供，隔离语义识别误差。", "contribution": "将机器人各运动连杆、障碍物和抓取后附着物建成可微Bernstein多项式SDF，以抓取模式切换受保护几何。对冻结π0.5的六维操作空间增量求最小改动CBF二次规划，保留原控制器和夹爪指令，无需重训策略。", "whyUseful": "复现必须公开任务角色排除、对象身份元数据、SDF拟合、接触搜索、松弛惩罚及控制缩放，分别验证几何误差与真实控制执行。正文没有实体机器人部署证据，不据LIBERO默认模型补填实机型号。", "analysisVerifiedAt": "2026-10-04T13:50:35.090120+00:00" }, { "id": "arxiv-2610.01102", "title": "MASkillBlender: Decentralized Whole-Body Coordination for Multi-Humanoid Loco-Manipulation via Skill Blending", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01102v1", "titleZh": "MASkillBlender：通过技能融合实现多人形机器人移动操作的分散式全身协调", "abstractZh": "MASkillBlender在冻结的单人形行走、伸手、下蹲技能之上，用共享局部策略输出技能目标和逐关节混合权重。MAPPO集中价值学习、分散执行；同构队员置换产生额外训练样本，并在同构博弈假设下分析其策略梯度合法性。下游主要任务奖励不代表低层技能无需密集奖励。\nIsaac Gym测试双机器人搬箱、推箱和三机器人避碰移动，主要为19自由度H1，另评21自由度G1。每任务两次独立训练，取最好策略做50回合；对照中央式SkillBlender和改编TeamHOI，另测延迟定位误差、技能/置换消融及MuJoCo跨模拟器迁移。", "summary": "MASkillBlender在冻结的单人形行走、伸手、下蹲技能之上，用共享局部策略输出技能目标和逐关节混合权重。MAPPO集中价值学习、分散执行；同构队员置换产生额外训练样本，并在同构博弈假设下分析其策略梯度合…", "experimentType": "sim", "robots": [ "Unitree H1（仿真）", "Unitree G1（仿真）" ], "tags": [ "人形机器人", "多智能体强化学习", "全身控制", "技能组合" ], "limitations": "所有验证均为仿真，包括名称含deployment的鲁棒性实验；没有H1/G1实机协作证据。取最优检查点及仅两次训练限制变异估计；技能库覆盖、同构假设及状态输入均限制实际泛化。Carry跨仿真前还增加随机化训练。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01102v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01102v1", "note": "5.1–5.2; Table 1：两训练试验，最好策略50回合；100/86/100%。" }, { "url": "https://arxiv.org/html/2610.01102v1", "note": "5.5：Carry跨模拟器前拓宽随机化再训练，Move直接迁移。" }, { "url": "https://arxiv.org/html/2610.01102v1", "note": "6; F.6：真机部署列未来挑战；Move含队友相对位置与通信延迟。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "人形机器人 / 多智能体强化学习", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "运动控制", "操作与抓取", "人形机器人", "多机器人", "强化学习" ], "robotFilters": [ "Unitree H1", "Unitree G1" ], "original": { "id": "arxiv-2610.01102", "originalSourceUrl": "https://arxiv.org/pdf/2610.01102v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01102v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01102v1", "sha256": "ef6791faecbb1ef9eaa8a608f9071c825cbf28e06d37cf11f5f13778b9e1695d", "bytes": 12049652, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:45.878812+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01102v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01102v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01102v1", "pages": 22, "sourceIdentityStatus": "title_match", "sourceTitle": "MASkillBlender: Decentralized Whole-Body Coordination for Multi-Humanoid Loco-Manipulation via Skill Blending", "sourceVersion": "2610.01102v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.519898+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01102v1" }, "originalAnalysis": { "id": "arxiv-2610.01102", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458235+00:00", "sourceUrl": "https://arxiv.org/html/2610.01102v1", "sectionsRead": [ "4.1–4.3", "5.1–5.6", "6 Discussion", "Appendix F.4–F.7", "Tables 1,4,5" ], "methodsZh": "MASkillBlender在冻结的单人形行走、伸手、下蹲技能之上，用共享局部策略输出技能目标和逐关节混合权重。MAPPO集中价值学习、分散执行；同构队员置换产生额外训练样本，并在同构博弈假设下分析其策略梯度合法性。下游主要任务奖励不代表低层技能无需密集奖励。", "experimentsZh": "Isaac Gym测试双机器人搬箱、推箱和三机器人避碰移动，主要为19自由度H1，另评21自由度G1。每任务两次独立训练，取最好策略做50回合；对照中央式SkillBlender和改编TeamHOI，另测延迟定位误差、技能/置换消融及MuJoCo跨模拟器迁移。", "resultsZh": "作者报告H1三任务成功率100%、86%、100%；去掉置换增强后搬箱降至6%，说明这一设置中增强作用显著。含定位和延迟扰动时成功率为92%、74%、98%，推箱仍最难。", "limitationsZh": "所有验证均为仿真，包括名称含deployment的鲁棒性实验；没有H1/G1实机协作证据。取最优检查点及仅两次训练限制变异估计；技能库覆盖、同构假设及状态输入均限制实际泛化。Carry跨仿真前还增加随机化训练。", "reproductionZh": "复现需取得各形态低层技能、技能选择与混合规则、全局critic/局部actor观测、任务成功阈值及置换实现。论文给出单5090训练约35–76小时，未计低层技能预训练；Move仍读取最近队友相对位置。", "experimentType": "sim", "robots": [ "Unitree H1（仿真）", "Unitree G1（仿真）" ], "evidenceNotes": [ { "section": "5.1–5.2; Table 1", "note": "两训练试验，最好策略50回合；100/86/100%。" }, { "section": "5.5", "note": "Carry跨模拟器前拓宽随机化再训练，Move直接迁移。" }, { "section": "6; F.6", "note": "真机部署列未来挑战；Move含队友相对位置与通信延迟。" } ], "corrections": [ "H1/G1型号必须标为仿真；Sim2Sim不能写成Sim2Real。" ], "sourceVersion": "2610.01102v1", "originalSha256": "ef6791faecbb1ef9eaa8a608f9071c825cbf28e06d37cf11f5f13778b9e1695d" }, "experimentNote": "Isaac Gym测试双机器人搬箱、推箱和三机器人避碰移动，主要为19自由度H1，另评21自由度G1。每任务两次独立训练，取最好策略做50回合；对照中央式SkillBlender和改编TeamHOI，另测延迟定位误差、技能/置换消融及MuJoCo跨模拟器迁移。", "contribution": "MASkillBlender在冻结的单人形行走、伸手、下蹲技能之上，用共享局部策略输出技能目标和逐关节混合权重。MAPPO集中价值学习、分散执行；同构队员置换产生额外训练样本，并在同构博弈假设下分析其策略梯度合法性。下游主要任务奖励不代表低层技能无需密集奖励。", "whyUseful": "复现需取得各形态低层技能、技能选择与混合规则、全局critic/局部actor观测、任务成功阈值及置换实现。论文给出单5090训练约35–76小时，未计低层技能预训练；Move仍读取最近队友相对位置。", "analysisVerifiedAt": "2026-10-04T13:45:34.458235+00:00" }, { "id": "arxiv-2610.01105", "title": "Extreme Length Generalization in a Compact Recurrent Architecture for One-Shot Exploration", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01105v1", "titleZh": "面向单次任务探索的紧凑循环架构及其极长序列泛化能力", "abstractZh": "FRANK将四种可学习时间常数的泄漏循环模块、键值记忆、无状态反射通路和抑制混合器结合，所有模块每步同时运行。算法任务版本50.7万参数，与GRU、RIMs、Mamba及删减结构按容量比较，重点检验训练长度外的稳定状态转移。\n主任务是随机数字的模10累加，训练长度5–20，测试最高200万步，十个种子、每极长单元十条流；另做copy/recall/sum权重损伤测试。MuJoCo PPO导航后部署ROSMASTER X3、树莓派5及RPLIDAR A1。", "summary": "FRANK将四种可学习时间常数的泄漏循环模块、键值记忆、无状态反射通路和抑制混合器结合，所有模块每步同时运行。算法任务版本50.7万参数，与GRU、RIMs、Mamba及删减结构按容量比较，重点检验训练长度外的…", "experimentType": "both", "robots": [ "ROSMASTER X3" ], "tags": [ "长时程任务", "循环架构", "记忆模型", "移动机器人" ], "limitations": "极长泛化来自简单有限状态任务，不证明机器人长任务可靠性；导航未给重复成功率或长期漂移试验。侧向连接6/10对3/10不显著，不能宣称已证实纠错；全局10%损伤下GRU明显更稳健。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01105v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01105v1", "note": "II–III：50.7万参数、十种子、六个完美泛化及统计。" }, { "url": "https://arxiv.org/html/2610.01105v1", "note": "IV–V：推翻部分侧连/鲁棒性解释；X3配置与定性范围。" }, { "url": "https://arxiv.org/pdf/2610.01105v1", "note": "II–III：50.7万参数、十种子、六个完美泛化及统计。" }, { "url": "https://arxiv.org/pdf/2610.01105v1", "note": "IV–V：推翻部分侧连/鲁棒性解释；X3配置与定性范围。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "长时程任务 / 循环架构", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "其他机器人研究" ], "robotFilters": [ "ROSMASTER X3" ], "original": { "id": "arxiv-2610.01105", "originalSourceUrl": "https://arxiv.org/pdf/2610.01105v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01105v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01105v1", "sha256": "f29179f46cccb463f01f4956158799b168dc7a25d9d1491ac4b574c114e9ef87", "bytes": 4379866, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:45.727933+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01105v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01105v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01105v1", "pages": 4, "sourceIdentityStatus": "title_match", "sourceTitle": "Extreme Length Generalization in a Compact Recurrent Architecture for One-Shot Exploration", "sourceVersion": "2610.01105v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.509553+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01105v1" }, "originalAnalysis": { "id": "arxiv-2610.01105", "sourceUrl": "https://arxiv.org/pdf/2610.01105v1", "sectionsRead": [ "II", "III", "IV", "V", "Figs 1–5" ], "methodsZh": "FRANK将四种可学习时间常数的泄漏循环模块、键值记忆、无状态反射通路和抑制混合器结合，所有模块每步同时运行。算法任务版本50.7万参数，与GRU、RIMs、Mamba及删减结构按容量比较，重点检验训练长度外的稳定状态转移。", "experimentsZh": "主任务是随机数字的模10累加，训练长度5–20，测试最高200万步，十个种子、每极长单元十条流；另做copy/recall/sum权重损伤测试。MuJoCo PPO导航后部署ROSMASTER X3、树莓派5及RPLIDAR A1。", "resultsZh": "十个种子中六个在所有长度保持100%，另有部分和失败解；100条超长流67条全程正确，基线500条无全正确。实机以13束激光、航点偏移和动作历史50Hz避障到点，仅给定性展示；导航网络扩宽到101万参数。", "limitationsZh": "极长泛化来自简单有限状态任务，不证明机器人长任务可靠性；导航未给重复成功率或长期漂移试验。侧向连接6/10对3/10不显著，不能宣称已证实纠错；全局10%损伤下GRU明显更稳健。", "reproductionZh": "保存每个训练种子的全长度表现，而非只展示最佳种子；长序列逐步核对精确状态，区分训练后损伤与从头删模块训练。导航应补定量路线、里程计漂移和安全停机测试，避免类比成行星或水下部署证据。", "experimentType": "both", "robots": [ "ROSMASTER X3" ], "corrections": [ "200万token泛化是模10序列任务；机器人仅定性导航，且部署模型为101万参数。" ], "evidenceNotes": [ { "note": "50.7万参数、十种子、六个完美泛化及统计。", "section": "II–III" }, { "note": "推翻部分侧连/鲁棒性解释；X3配置与定性范围。", "section": "IV–V" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2610.01105v1", "originalSha256": "f29179f46cccb463f01f4956158799b168dc7a25d9d1491ac4b574c114e9ef87" }, "experimentNote": "主任务是随机数字的模10累加，训练长度5–20，测试最高200万步，十个种子、每极长单元十条流；另做copy/recall/sum权重损伤测试。MuJoCo PPO导航后部署ROSMASTER X3、树莓派5及RPLIDAR A1。", "contribution": "FRANK将四种可学习时间常数的泄漏循环模块、键值记忆、无状态反射通路和抑制混合器结合，所有模块每步同时运行。算法任务版本50.7万参数，与GRU、RIMs、Mamba及删减结构按容量比较，重点检验训练长度外的稳定状态转移。", "whyUseful": "保存每个训练种子的全长度表现，而非只展示最佳种子；长序列逐步核对精确状态，区分训练后损伤与从头删模块训练。导航应补定量路线、里程计漂移和安全停机测试，避免类比成行星或水下部署证据。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2610.01162", "title": "PhysicsLENS: Diagnosing Physical Property Blindness in Video Generation Models", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01162v1", "titleZh": "PhysicsLENS：诊断视频生成模型对物理属性的忽视", "abstractZh": "用相同起始图和任务构造可见条件与文本指定隐藏物性配对，分别人工评分物理合理性、任务完成和隐藏属性遵循。诊断器把相机补偿轨迹、光流和物体消失等信号，与针对具体物理错误的VLM问题组合，避免仅询问整体真实感。\n80幅起始图来自11个机器人数据集，形成110提示，四种视频模型共生成439段，隐藏属性119段。每提示模型仅一次生成；自动评测采用按场景分组五折验证，十个VLM及不同取帧重复，人工标签不含时间定位。", "summary": "用相同起始图和任务构造可见条件与文本指定隐藏物性配对，分别人工评分物理合理性、任务完成和隐藏属性遵循。诊断器把相机补偿轨迹、光流和物体消失等信号，与针对具体物理错误的VLM问题组合，避免仅询问整体真实感。", "experimentType": "data", "robots": [ "未核实" ], "tags": [ "视频世界模型", "物理推理", "评测基准", "生成模型" ], "limitations": "结果针对这批生成视频，不估计不同生成种子的方差；覆盖模型、物性及场景有限。自动器能排序并不证明逐例识别了隐藏属性；错误归因较弱，也未评估报告的时间定位，更不代表机器人实机能力。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01162v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/pdf/2610.01162v1", "note": "§3.2–3.4：439段、单次生成、三类标签及自动信号。" }, { "url": "https://arxiv.org/pdf/2610.01162v1", "note": "§4.1–4.4 Tables 2–5：34/47与15/28交叉失败、0.72/0.81 AUC及不显著配对差异。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "视频世界模型 / 物理推理", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "世界模型", "数据集与基准", "安全与评估" ], "robotFilters": [], "original": { "id": "arxiv-2610.01162", "originalSourceUrl": "https://arxiv.org/pdf/2610.01162v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01162v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01162v1", "sha256": "f2d63c19df085c547f9f4609ca4b09e7f3651f0cca749e07c2f1b26ea702cc06", "bytes": 6754311, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:41.240701+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01162v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01162v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01162v1", "pages": 21, "sourceIdentityStatus": "title_match", "sourceTitle": "PhysicsLENS: Diagnosing Physical Property Blindness in Video Generation Models", "sourceVersion": "2610.01162v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.505448+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01162v1" }, "originalAnalysis": { "id": "arxiv-2610.01162", "sourceUrl": "https://arxiv.org/pdf/2610.01162v1", "sectionsRead": [ "§3.1–3.4", "§4.1–4.4", "Tables 2–5", "§5" ], "methodsZh": "用相同起始图和任务构造可见条件与文本指定隐藏物性配对，分别人工评分物理合理性、任务完成和隐藏属性遵循。诊断器把相机补偿轨迹、光流和物体消失等信号，与针对具体物理错误的VLM问题组合，避免仅询问整体真实感。", "experimentsZh": "80幅起始图来自11个机器人数据集，形成110提示，四种视频模型共生成439段，隐藏属性119段。每提示模型仅一次生成；自动评测采用按场景分组五折验证，十个VLM及不同取帧重复，人工标签不含时间定位。", "resultsZh": "作者发现隐藏条件中47段被评为合理，其中34段不遵循指定物性；反过来28段遵循属性的视频有15段物理不合理。完整评测器合理性/任务完成AUC为0.72/0.81。隐藏条件合理性平均下降0.28分，但p=0.085，不能称显著下降。", "limitationsZh": "结果针对这批生成视频，不估计不同生成种子的方差；覆盖模型、物性及场景有限。自动器能排序并不证明逐例识别了隐藏属性；错误归因较弱，也未评估报告的时间定位，更不代表机器人实机能力。", "reproductionZh": "必须保持配对起始图和任务、按场景划分交叉验证，分别报告三类评分；不要以任务完成替代物理合理性，或把机器人原始图像算作部署试验。本次分析未重新生成视频。", "experimentType": "data", "robots": [], "evidenceNotes": [ { "section/page": "§3.2–3.4", "note": "439段、单次生成、三类标签及自动信号。" }, { "section/page": "§4.1–4.4 Tables 2–5", "note": "34/47与15/28交叉失败、0.72/0.81 AUC及不显著配对差异。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "80幅起始图来自11个机器人数据集，形成110提示，四种视频模型共生成439段，隐藏属性119段。每提示模型仅一次生成；自动评测采用按场景分组五折验证，十个VLM及不同取帧重复，人工标签不含时间定位。", "contribution": "用相同起始图和任务构造可见条件与文本指定隐藏物性配对，分别人工评分物理合理性、任务完成和隐藏属性遵循。诊断器把相机补偿轨迹、光流和物体消失等信号，与针对具体物理错误的VLM问题组合，避免仅询问整体真实感。", "whyUseful": "必须保持配对起始图和任务、按场景划分交叉验证，分别报告三类评分；不要以任务完成替代物理合理性，或把机器人原始图像算作部署试验。本次分析未重新生成视频。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2610.01171", "title": "Experience-Based Feasibility-Aware Generative Adversarial Imitation from Observation under Embodiment Mismatch", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01171v1", "titleZh": "形态不匹配条件下，基于经验与可行性感知的观测式生成对抗模仿学习", "abstractZh": "EF-GAIfO只用状态转移示范，以机器人自身探索训练重建模型；重建误差转换成可行性权重，对专家转移的判别器损失降权，PPO策略、判别器与重建模型共同迭代。它估计与经验相似的运动可行性，不需要专家动作标签。\n二维点质量分别混合超速/可行速度及不同路径，20随机种子；Go2任务从30条人类无机器人示范学接近、抓取、抬升，高层控制速度、俯仰和夹爪，低层步态预训练。仿真每设置5策略各100次，真机四目标设置各10次。", "summary": "EF-GAIfO只用状态转移示范，以机器人自身探索训练重建模型；重建误差转换成可行性权重，对专家转移的判别器损失降权，PPO策略、判别器与重建模型共同迭代。它估计与经验相似的运动可行性，不需要专家动作标签。", "experimentType": "both", "robots": [ "Unitree Go2" ], "tags": [ "观测式模仿学习", "形态迁移", "四足机器人", "可行性学习" ], "limitations": "误差小仅说明接近已有经验，不是物理可行性证明；早期探索不足时可能抑制可行但未见的动作。目标随机化只在邻近区域，真实样本较少；估计器未充分纳入地形、障碍和接触环境。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01171v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01171v1", "note": "§IV-A：以自采经验重建误差估计可行性。" }, { "url": "https://arxiv.org/html/2610.01171v1", "note": "Tables III–IV：仿真5种子×100，真机每设置10次；分开报告。" }, { "url": "https://arxiv.org/html/2610.01171v1", "note": "§VI：环境依赖可行性尚未建模。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "观测式模仿学习 / 形态迁移", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "运动控制", "模仿学习" ], "robotFilters": [ "Unitree Go2" ], "original": { "id": "arxiv-2610.01171", "originalSourceUrl": "https://arxiv.org/pdf/2610.01171v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01171v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01171v1", "sha256": "2c5f073c7d74083ec3f9275d2a2740c56fa977f03a942473912a2511ef5b0172", "bytes": 2995207, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:33.964594+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01171v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01171v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01171v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "Experience-Based Feasibility-Aware Generative Adversarial Imitation from Observation under Embodiment Mismatch", "sourceVersion": "2610.01171v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.498884+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01171v1" }, "originalAnalysis": { "id": "arxiv-2610.01171", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090117+00:00", "sourceUrl": "https://arxiv.org/html/2610.01171v1", "sectionsRead": [ "§IV", "§V-A–V-B", "Tables I–IV", "§VI" ], "methodsZh": "EF-GAIfO只用状态转移示范，以机器人自身探索训练重建模型；重建误差转换成可行性权重，对专家转移的判别器损失降权，PPO策略、判别器与重建模型共同迭代。它估计与经验相似的运动可行性，不需要专家动作标签。", "experimentsZh": "二维点质量分别混合超速/可行速度及不同路径，20随机种子；Go2任务从30条人类无机器人示范学接近、抓取、抬升，高层控制速度、俯仰和夹爪，低层步态预训练。仿真每设置5策略各100次，真机四目标设置各10次。", "resultsZh": "作者报告仿真随机目标成功85.8±7.4%，WGAIfO/GAIfO为26.4±44.1%/35.4±18.1%。真实Go2四设置成功70–80%，对照10–30%，说明指定抓取任务中筛除不合适示范具有帮助。", "limitationsZh": "误差小仅说明接近已有经验，不是物理可行性证明；早期探索不足时可能抑制可行但未见的动作。目标随机化只在邻近区域，真实样本较少；估计器未充分纳入地形、障碍和接触环境。", "reproductionZh": "复现需固定示范可行/不可行混合、阈值分位调度和经验缓冲，单独核验低层步态及夹爪安装。本体型号有明确Go2证据，不能把人类采集状态直接当机器人动作示范；当前未运行实现。", "experimentType": "both", "robots": [ "Unitree Go2" ], "evidenceNotes": [ { "section/page": "§IV-A", "note": "以自采经验重建误差估计可行性。" }, { "section/page": "Tables III–IV", "note": "仿真5种子×100，真机每设置10次；分开报告。" }, { "section/page": "§VI", "note": "环境依赖可行性尚未建模。" } ] }, "experimentNote": "二维点质量分别混合超速/可行速度及不同路径，20随机种子；Go2任务从30条人类无机器人示范学接近、抓取、抬升，高层控制速度、俯仰和夹爪，低层步态预训练。仿真每设置5策略各100次，真机四目标设置各10次。", "contribution": "EF-GAIfO只用状态转移示范，以机器人自身探索训练重建模型；重建误差转换成可行性权重，对专家转移的判别器损失降权，PPO策略、判别器与重建模型共同迭代。它估计与经验相似的运动可行性，不需要专家动作标签。", "whyUseful": "复现需固定示范可行/不可行混合、阈值分位调度和经验缓冲，单独核验低层步态及夹爪安装。本体型号有明确Go2证据，不能把人类采集状态直接当机器人动作示范；当前未运行实现。", "analysisVerifiedAt": "2026-10-04T13:50:35.090117+00:00" }, { "id": "arxiv-2610.01178", "title": "Recova: Agent-Guided Failure Recovery for Autonomous Robotic Manipulation", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01178v1", "titleZh": "Recova：智能体引导的自主机器人操作失败恢复", "abstractZh": "Recova把完成任务与恢复可工作场景拆成两个策略。编码智能体先用真实影像、轨迹和标定搭MuJoCo数字孪生，在失败状态试验纠错程序；上线后视觉语言监控检查进展、选择已登记恢复技能并验证结果，不适用时收集人类示范，按任务/恢复用途分别聚合微调。\n仿真以MolmoAct2配程序恢复，LIBERO-Pro六设置每任务50初态及MolmoSpaces四类，与已发表基线对照。实机为四套I2RT YAM双臂工作站，π0.5任务/恢复策略，四任务每配置20次；单人可监督四站，独立记录数据收集轮次的人工接管。", "summary": "Recova把完成任务与恢复可工作场景拆成两个策略。编码智能体先用真实影像、轨迹和标定搭MuJoCo数字孪生，在失败状态试验纠错程序；上线后视觉语言监控检查进展、选择已登记恢复技能并验证结果，不适用时收集人类示…", "experimentType": "both", "robots": [ "I2RT YAM（双臂工作站）" ], "tags": [ "失败恢复", "数字孪生", "机器人操作", "人在回路学习" ], "limitations": "仿真基线引用既有论文，并非全部同环境重跑；真机成功由操作员判定。未知技能仍需人示范，监控回复有数秒延迟，失败片段被排除出监督训练。数字孪生接触误差和技能注册仍限制自动扩展。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01178v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01178v1", "note": "4.1：Recova自评，基线数值来自既有发表结果。" }, { "url": "https://arxiv.org/html/2610.01178v1", "note": "4.2; Table 3：四实机任务每配置20次；23.8/77.5/87.5%。" }, { "url": "https://arxiv.org/html/2610.01178v1", "note": "4.2 DAgger rounds; C.4：零干预是单任务七回合；实机评价操作员判定。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "失败恢复 / 数字孪生", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "操作与抓取" ], "robotFilters": [ "YAM" ], "original": { "id": "arxiv-2610.01178", "originalSourceUrl": "https://arxiv.org/pdf/2610.01178v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01178v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01178v1", "sha256": "9e97f8d4a2ad86a447748a9dd968b97a18230b42a884c5a68610b6aaeb1e7a64", "bytes": 6838410, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:34.079063+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01178v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01178v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01178v1", "pages": 23, "sourceIdentityStatus": "title_match", "sourceTitle": "Recova: Agent-Guided Failure Recovery for Autonomous Robotic Manipulation", "sourceVersion": "2610.01178v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.495306+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01178v1" }, "originalAnalysis": { "id": "arxiv-2610.01178", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458233+00:00", "sourceUrl": "https://arxiv.org/html/2610.01178v1", "sectionsRead": [ "3.1–3.4", "4.1–4.2", "Tables 1–3", "Appendix A", "Appendix B", "Appendix C.1–C.5" ], "methodsZh": "Recova把完成任务与恢复可工作场景拆成两个策略。编码智能体先用真实影像、轨迹和标定搭MuJoCo数字孪生，在失败状态试验纠错程序；上线后视觉语言监控检查进展、选择已登记恢复技能并验证结果，不适用时收集人类示范，按任务/恢复用途分别聚合微调。", "experimentsZh": "仿真以MolmoAct2配程序恢复，LIBERO-Pro六设置每任务50初态及MolmoSpaces四类，与已发表基线对照。实机为四套I2RT YAM双臂工作站，π0.5任务/恢复策略，四任务每配置20次；单人可监督四站，独立记录数据收集轮次的人工接管。", "resultsZh": "作者报告仿真均值78.8%和64.9%；实机基策略23.8%，DAgger后77.5%，加恢复87.5%。摸麻将任务第四轮0%接管仅来自七个收集回合，第一轮为八回合中七次；不是跨四任务长期无人干预保证。", "limitationsZh": "仿真基线引用既有论文，并非全部同环境重跑；真机成功由操作员判定。未知技能仍需人示范，监控回复有数秒延迟，失败片段被排除出监督训练。数字孪生接触误差和技能注册仍限制自动扩展。", "reproductionZh": "附录规定30Hz、每任务尝试三分钟、每恢复50秒，完成/干预检查分别20/10秒；微调使用八L40、每阶段2K–5K步。复现应保留策略来源、视频判决、人工重标及回合边界，以免把收集期人工完成计入自主测试。", "experimentType": "both", "robots": [ "I2RT YAM（双臂工作站）" ], "evidenceNotes": [ { "section": "4.1", "note": "Recova自评，基线数值来自既有发表结果。" }, { "section": "4.2; Table 3", "note": "四实机任务每配置20次；23.8/77.5/87.5%。" }, { "section": "4.2 DAgger rounds; C.4", "note": "零干预是单任务七回合；实机评价操作员判定。" } ], "corrections": [ "补充硬件I2RT YAM；0%干预仅限一项任务第四收集轮七回合。" ], "sourceVersion": "2610.01178v1", "originalSha256": "9e97f8d4a2ad86a447748a9dd968b97a18230b42a884c5a68610b6aaeb1e7a64" }, "experimentNote": "仿真以MolmoAct2配程序恢复，LIBERO-Pro六设置每任务50初态及MolmoSpaces四类，与已发表基线对照。实机为四套I2RT YAM双臂工作站，π0.5任务/恢复策略，四任务每配置20次；单人可监督四站，独立记录数据收集轮次的人工接管。", "contribution": "Recova把完成任务与恢复可工作场景拆成两个策略。编码智能体先用真实影像、轨迹和标定搭MuJoCo数字孪生，在失败状态试验纠错程序；上线后视觉语言监控检查进展、选择已登记恢复技能并验证结果，不适用时收集人类示范，按任务/恢复用途分别聚合微调。", "whyUseful": "附录规定30Hz、每任务尝试三分钟、每恢复50秒，完成/干预检查分别20/10秒；微调使用八L40、每阶段2K–5K步。复现应保留策略来源、视频判决、人工重标及回合边界，以免把收集期人工完成计入自主测试。", "analysisVerifiedAt": "2026-10-04T13:45:34.458233+00:00" }, { "id": "arxiv-2610.01219", "title": "EIDA: Execution-Interface Dynamics Adaptation for Real-to-Sim-to-Real Robot Navigation", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01219v1", "titleZh": "EIDA：面向真实—仿真—真实机器人导航的执行接口动力学适配", "abstractZh": "EIDA专门拟合高层速度命令经过现有底层控制器和估计器后的执行接口：MLP预测机体坐标位姿增量，独立ARX预测策略实际收到的速度估计。将两者植入快速二维模拟器，以带速度历史的SAC学导航；部署时两拟合模型都不在线运行。\nJackal从目标Gazebo环境采样、在BARN100世界每配置300次评估；真实Go2从FAST-LIO2位姿和Sport速度反馈采数据，10Hz部署于Jetson Orin NX，配MID-360，两个人造草坪障碍场景各10次。", "summary": "EIDA专门拟合高层速度命令经过现有底层控制器和估计器后的执行接口：MLP预测机体坐标位姿增量，独立ARX预测策略实际收到的速度估计。将两者植入快速二维模拟器，以带速度历史的SAC学导航；部署时两拟合模型都不在…", "experimentType": "both", "robots": [ "Clearpath Jackal（仿真）", "Unitree Go2（实机）" ], "tags": [ "自主导航", "真实仿真迁移", "系统辨识", "四足机器人" ], "limitations": "实机仅二场景20个方法试验；动态行人是另行定性展示。路径和时间只对成功试验统计，不能用其均值忽略失败；模型必须覆盖目标平台接口，不能据此推断跨平台零样本迁移。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01219v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01219v1", "note": "IV; V-A：双模型执行接口及部署时不运行。" }, { "url": "https://arxiv.org/html/2610.01219v1", "note": "V-C Table I; V-E Table III：Jackal仿真、Go2实机范围、成功分母与阈值。" }, { "url": "https://arxiv.org/pdf/2610.01219v1", "note": "IV; V-A：双模型执行接口及部署时不运行。" }, { "url": "https://arxiv.org/pdf/2610.01219v1", "note": "V-C Table I; V-E Table III：Jackal仿真、Go2实机范围、成功分母与阈值。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "自主导航 / 真实仿真迁移", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "运动控制", "导航与建图" ], "robotFilters": [ "Clearpath Jackal", "Unitree Go2" ], "original": { "id": "arxiv-2610.01219", "originalSourceUrl": "https://arxiv.org/pdf/2610.01219v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01219v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01219v1", "sha256": "67bb3ec388afa7ec9dd899f5245f38a7383ae7128e7e4eb9f58f470649bad7d8", "bytes": 5722978, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:24.159287+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01219v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01219v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01219v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "EIDA: Execution-Interface Dynamics Adaptation for Real-to-Sim-to-Real Robot Navigation", "sourceVersion": "2610.01219v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.489739+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01219v1" }, "originalAnalysis": { "id": "arxiv-2610.01219", "sourceUrl": "https://arxiv.org/pdf/2610.01219v1", "sectionsRead": [ "III", "IV-A–B", "V-A–E", "VI", "Tables I–III" ], "methodsZh": "EIDA专门拟合高层速度命令经过现有底层控制器和估计器后的执行接口：MLP预测机体坐标位姿增量，独立ARX预测策略实际收到的速度估计。将两者植入快速二维模拟器，以带速度历史的SAC学导航；部署时两拟合模型都不在线运行。", "experimentsZh": "Jackal从目标Gazebo环境采样、在BARN100世界每配置300次评估；真实Go2从FAST-LIO2位姿和Sport速度反馈采数据，10Hz部署于Jetson Orin NX，配MID-360，两个人造草坪障碍场景各10次。", "resultsZh": "有全局路径时BARN成功89.33%，基线80.78%；Go2两场景均10/10无碰撞到达，基线1/10及3/10。五秒预测在完整验证集改善，但Go2部署命令范围内位置精度不始终更好；消融说明几何变化和估计反馈不能直接互换。", "limitationsZh": "实机仅二场景20个方法试验；动态行人是另行定性展示。路径和时间只对成功试验统计，不能用其均值忽略失败；模型必须覆盖目标平台接口，不能据此推断跨平台零样本迁移。", "reproductionZh": "先按轨迹分组划分辨识数据，分别检查命令历史、速度历史和反馈来源；保持底层控制器及状态估计器版本一致。模拟成功门槛1米/100秒与实机0.3米/30秒不同，报告时不要混并。", "experimentType": "both", "robots": [ "Clearpath Jackal（仿真）", "Unitree Go2（实机）" ], "corrections": [ "Jackal结果为sim-to-sim，真实部署仅Go2；拟合动态模型只用于训练环境。" ], "evidenceNotes": [ { "note": "双模型执行接口及部署时不运行。", "section": "IV; V-A" }, { "note": "Jackal仿真、Go2实机范围、成功分母与阈值。", "section": "V-C Table I; V-E Table III" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2610.01219v1", "originalSha256": "67bb3ec388afa7ec9dd899f5245f38a7383ae7128e7e4eb9f58f470649bad7d8" }, "experimentNote": "Jackal从目标Gazebo环境采样、在BARN100世界每配置300次评估；真实Go2从FAST-LIO2位姿和Sport速度反馈采数据，10Hz部署于Jetson Orin NX，配MID-360，两个人造草坪障碍场景各10次。", "contribution": "EIDA专门拟合高层速度命令经过现有底层控制器和估计器后的执行接口：MLP预测机体坐标位姿增量，独立ARX预测策略实际收到的速度估计。将两者植入快速二维模拟器，以带速度历史的SAC学导航；部署时两拟合模型都不在线运行。", "whyUseful": "先按轨迹分组划分辨识数据，分别检查命令历史、速度历史和反馈来源；保持底层控制器及状态估计器版本一致。模拟成功门槛1米/100秒与实机0.3米/30秒不同，报告时不要混并。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2610.01224", "title": "Supervise What Decides Success: Criterion-Aligned Auxiliary Losses for Latent World-Model Planning", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01224v1", "titleZh": "监督决定成功的量：用于潜空间世界模型规划的成功准则对齐辅助损失", "abstractZh": "给潜在世界模型编码器和预测器共享线性读出头，只监督任务成功判据涉及的位置等物理量；原有预测损失保持不变。训练后移除辅助头，CEM仍按潜在终态与目标图像的距离规划，使成功所需精度通过表征进入候选排序。\n在LeWM五个仿真任务中逐项加入手、物体或关节监督，以相同初始/目标对比较2000次试验；DINO-WM型模型使用1000对。报告配对McNemar检验，并将只需到达和必须搬动物体的试验分层，检查整体平均掩盖的反向影响。", "summary": "给潜在世界模型编码器和预测器共享线性读出头，只监督任务成功判据涉及的位置等物理量；原有预测损失保持不变。训练后移除辅助头，CEM仍按潜在终态与目标图像的距离规划，使成功所需精度通过表征进入候选排序。", "experimentType": "sim", "robots": [ "未核实" ], "tags": [ "世界模型", "潜空间规划", "表征学习", "辅助监督" ], "limitations": "多数条件只有一个训练种子，2000对测试不能替代训练随机性评估；冻结编码器复核主要限stack，冻结V-JEPA 2未复现改善。依赖已知且可从图像辨识的成功判据，未验证真实机器人或接触类逻辑判据。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01224v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/pdf/2610.01224v1", "note": "§2.3：共享读出头仅训练使用，测试仍采用潜在距离。" }, { "url": "https://arxiv.org/pdf/2610.01224v1", "note": "Tables 1–3; §5：2000/1000配对协议、增益与单种子、V-JEPA负结果。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "世界模型 / 潜空间规划", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "世界模型" ], "robotFilters": [], "original": { "id": "arxiv-2610.01224", "originalSourceUrl": "https://arxiv.org/pdf/2610.01224v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01224v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01224v1", "sha256": "3aa8ed58541abbc8a6bd56f12fa4b8bbe1ce00e3dab4c8c53cfe9e60c30edfa7", "bytes": 1823996, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:17.752495+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01224v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01224v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01224v1", "pages": 21, "sourceIdentityStatus": "title_match", "sourceTitle": "Supervise What Decides Success: Criterion-Aligned Auxiliary Losses for Latent World-Model Planning", "sourceVersion": "2610.01224v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.484202+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01224v1" }, "originalAnalysis": { "id": "arxiv-2610.01224", "sourceUrl": "https://arxiv.org/pdf/2610.01224v1", "sectionsRead": [ "§2", "§3 Tables 1–3", "§5", "Appendix A" ], "methodsZh": "给潜在世界模型编码器和预测器共享线性读出头，只监督任务成功判据涉及的位置等物理量；原有预测损失保持不变。训练后移除辅助头，CEM仍按潜在终态与目标图像的距离规划，使成功所需精度通过表征进入候选排序。", "experimentsZh": "在LeWM五个仿真任务中逐项加入手、物体或关节监督，以相同初始/目标对比较2000次试验；DINO-WM型模型使用1000对。报告配对McNemar检验，并将只需到达和必须搬动物体的试验分层，检查整体平均掩盖的反向影响。", "resultsZh": "作者报告stack同时监督手和物体，相对30.1%基线提升5.05个百分点；coffee仅监督物体提升3.90点，全部监督却为不显著的负0.60点。DINOv2-S模型stack监督手提高7.10点。线性读出变准并不自动保证控制成功。", "limitationsZh": "多数条件只有一个训练种子，2000对测试不能替代训练随机性评估；冻结编码器复核主要限stack，冻结V-JEPA 2未复现改善。依赖已知且可从图像辨识的成功判据，未验证真实机器人或接触类逻辑判据。", "reproductionZh": "应固定CEM预算、起终点、容差及辅助权重，把所有成功条件一并列明；不仅测回归误差，还测候选排序和任务成功。需保留到达/搬运分层及负结果，本次未运行训练。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "§2.3", "note": "共享读出头仅训练使用，测试仍采用潜在距离。" }, { "section/page": "Tables 1–3; §5", "note": "2000/1000配对协议、增益与单种子、V-JEPA负结果。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "在LeWM五个仿真任务中逐项加入手、物体或关节监督，以相同初始/目标对比较2000次试验；DINO-WM型模型使用1000对。报告配对McNemar检验，并将只需到达和必须搬动物体的试验分层，检查整体平均掩盖的反向影响。", "contribution": "给潜在世界模型编码器和预测器共享线性读出头，只监督任务成功判据涉及的位置等物理量；原有预测损失保持不变。训练后移除辅助头，CEM仍按潜在终态与目标图像的距离规划，使成功所需精度通过表征进入候选排序。", "whyUseful": "应固定CEM预算、起终点、容差及辅助权重，把所有成功条件一并列明；不仅测回归误差，还测候选排序和任务成功。需保留到达/搬运分层及负结果，本次未运行训练。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2610.01258", "title": "ColoACT: Multi-Cue Action Chunking for Smooth Autonomous Colon Navigation on a Self-Propelled Endoscopic Robot", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01258v1", "titleZh": "ColoACT：面向自推进内镜机器人的多线索动作分块方法，实现平滑自主结肠导航", "abstractZh": "单目RGB由Depth Anything V2产生相对深度，再用梯度生成伪高程，组成五通道输入；改造ResNet-18和ACT的CVAE预测16步线/角速度块，以重叠块均值平滑命令。定制BGER履带与锥齿轮结构降低机械抖动，视频经系绳传到主机。\n214条离体猪结肠遥操作轨迹训练100轮，A100训练、RTX4060以15Hz部署。在四条训练外约60厘米结肠进行48次直线、40次弯曲试验，另对直角、双弯、三弯各测10次；对照相同输入的单步BC和输入消融。", "summary": "单目RGB由Depth Anything V2产生相对深度，再用梯度生成伪高程，组成五通道输入；改造ResNet-18和ACT的CVAE预测16步线/角速度块，以重叠块均值平滑命令。定制BGER履带与锥齿轮结构…", "experimentType": "real", "robots": [ "BGER（定制自推进内窥镜机器人）" ], "tags": [ "医疗机器人", "离体实验", "模仿学习", "自主导航" ], "limitations": "全部是离体实组织，未有活体/患者验证；深度是估计相对深度，伪高程并非直接测量。命令平滑不证明组织安全，缺少接触力传感，蠕动和润滑变化仍待验证。注意复杂场景样本量很小。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01258v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01258v1", "note": "§IV.1–IV.2：214训练轨迹；4新结肠；48直线/40弯曲试验。" }, { "url": "https://arxiv.org/html/2610.01258v1", "note": "§IV.2：三弯3/10；不得只写可通过而隐去失败比例。" }, { "url": "https://arxiv.org/html/2610.01258v1", "note": "§V：活体验证和力传感为未来工作。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "医疗机器人 / 离体实验", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "导航与建图", "模仿学习" ], "robotFilters": [], "original": { "id": "arxiv-2610.01258", "originalSourceUrl": "https://arxiv.org/pdf/2610.01258v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01258v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01258v1", "sha256": "dc41ef930d411340bf530499d13732c53829aa55262a7f463b67356f61c38f0b", "bytes": 5380356, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:17.804016+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01258v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01258v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01258v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "ColoACT: Multi-Cue Action Chunking for Smooth Autonomous Colon Navigation on a Self-Propelled Endoscopic Robot", "sourceVersion": "2610.01258v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.480962+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01258v1" }, "originalAnalysis": { "id": "arxiv-2610.01258", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090115+00:00", "sourceUrl": "https://arxiv.org/html/2610.01258v1", "sectionsRead": [ "§III", "§IV.1–IV.4", "Tables I–III", "§V" ], "methodsZh": "单目RGB由Depth Anything V2产生相对深度，再用梯度生成伪高程，组成五通道输入；改造ResNet-18和ACT的CVAE预测16步线/角速度块，以重叠块均值平滑命令。定制BGER履带与锥齿轮结构降低机械抖动，视频经系绳传到主机。", "experimentsZh": "214条离体猪结肠遥操作轨迹训练100轮，A100训练、RTX4060以15Hz部署。在四条训练外约60厘米结肠进行48次直线、40次弯曲试验，另对直角、双弯、三弯各测10次；对照相同输入的单步BC和输入消融。", "resultsZh": "作者报告直线85.4%、弯曲72.5%，BC为45.8%/57.5%；命令角加加速度约由60降至不足5。复杂路径为7/10、6/10和3/10，三弯仅显示有限可行性；43样本离线ADE/FDE另行统计。", "limitationsZh": "全部是离体实组织，未有活体/患者验证；深度是估计相对深度，伪高程并非直接测量。命令平滑不证明组织安全，缺少接触力传感，蠕动和润滑变化仍待验证。注意复杂场景样本量很小。", "reproductionZh": "复现须公开按结肠个体划分的训练/测试集、动作归一化、平滑窗口与BC训练条件，同时将离线预测误差和闭环通过率分开。原文计划加入力传感并做活体验证，本轮未独立测试硬件或训练。", "experimentType": "real", "robots": [ "BGER（定制自推进内窥镜机器人）" ], "evidenceNotes": [ { "section/page": "§IV.1–IV.2", "note": "214训练轨迹；4新结肠；48直线/40弯曲试验。" }, { "section/page": "§IV.2", "note": "三弯3/10；不得只写可通过而隐去失败比例。" }, { "section/page": "§V", "note": "活体验证和力传感为未来工作。" } ] }, "experimentNote": "214条离体猪结肠遥操作轨迹训练100轮，A100训练、RTX4060以15Hz部署。在四条训练外约60厘米结肠进行48次直线、40次弯曲试验，另对直角、双弯、三弯各测10次；对照相同输入的单步BC和输入消融。", "contribution": "单目RGB由Depth Anything V2产生相对深度，再用梯度生成伪高程，组成五通道输入；改造ResNet-18和ACT的CVAE预测16步线/角速度块，以重叠块均值平滑命令。定制BGER履带与锥齿轮结构降低机械抖动，视频经系绳传到主机。", "whyUseful": "复现须公开按结肠个体划分的训练/测试集、动作归一化、平滑窗口与BC训练条件，同时将离线预测误差和闭环通过率分开。原文计划加入力传感并做活体验证，本轮未独立测试硬件或训练。", "analysisVerifiedAt": "2026-10-04T13:50:35.090115+00:00" }, { "id": "arxiv-2610.01260", "title": "PROMO: Preference-conditioned Multi-Objective Reinforcement Learning for Quadrupedal Robots", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01260v1", "titleZh": "PROMO：面向四足机器人的偏好条件化多目标强化学习", "abstractZh": "PROMO把速度跟踪、身体稳定、能耗效率分成可调语义目标，将步态与接触正则保持为固定先验。偏好向量同时输入策略和历史编码，三头价值网络分别学习语义回报；特权编码在训练中对齐可部署历史表示，部署只保留本体观测、速度估计与单一策略。\nIsaacLab中4096个Go2平地环境训练，随机化摩擦、惯性、质心和扰动。模拟密集测试100偏好；实机Go2在Vicon场地使用同一检查点，四种偏好乘三种指令条件各五次，策略50Hz、底层200Hz。每段试验偏好固定，只有重放条件严格共享指令轨迹。", "summary": "PROMO把速度跟踪、身体稳定、能耗效率分成可调语义目标，将步态与接触正则保持为固定先验。偏好向量同时输入策略和历史编码，三头价值网络分别学习语义回报；特权编码在训练中对齐可部署历史表示，部署只保留本体观测、速…", "experimentType": "both", "robots": [ "Unitree Go2" ], "tags": [ "四足机器人", "多目标强化学习", "偏好条件化", "运动控制" ], "limitations": "跟踪位置更好可伴随速度RMSE或能耗更差；慢速稳定偏好的速度误差为2.203米/秒。五次/条件不能支持密集真机可控性或统计显著性，非支配样本不是全球Pareto最优证明，手动扰动恢复仅定性。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01260v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01260v1", "note": "III-A：所有报告评测每回合偏好固定，非支配集合为有限采样近似。" }, { "url": "https://arxiv.org/html/2610.01260v1", "note": "V-A; Table III：四偏好×三条件×五试验；实机Go2。" }, { "url": "https://arxiv.org/html/2610.01260v1", "note": "V-B; VI-C：收益存在交叉目标代价；实机不主张统计显著性。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "四足机器人 / 多目标强化学习", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "运动控制", "强化学习" ], "robotFilters": [ "Unitree Go2" ], "original": { "id": "arxiv-2610.01260", "originalSourceUrl": "https://arxiv.org/pdf/2610.01260v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01260v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01260v1", "sha256": "91068a3909fad60606b289d8e3432d205e0dc13440c32010f4882c2f923c0666", "bytes": 8146975, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:11.362912+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01260v1", "license": "http://creativecommons.org/licenses/by-sa/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01260v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01260v1", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "PROMO: Preference-conditioned Multi-Objective Reinforcement Learning for Quadrupedal Robots", "sourceVersion": "2610.01260v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.475788+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01260v1" }, "originalAnalysis": { "id": "arxiv-2610.01260", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458232+00:00", "sourceUrl": "https://arxiv.org/html/2610.01260v1", "sectionsRead": [ "III-A–D", "IV-A–E", "V-A–B", "VI-C", "Table III", "Appendix A–B" ], "methodsZh": "PROMO把速度跟踪、身体稳定、能耗效率分成可调语义目标，将步态与接触正则保持为固定先验。偏好向量同时输入策略和历史编码，三头价值网络分别学习语义回报；特权编码在训练中对齐可部署历史表示，部署只保留本体观测、速度估计与单一策略。", "experimentsZh": "IsaacLab中4096个Go2平地环境训练，随机化摩擦、惯性、质心和扰动。模拟密集测试100偏好；实机Go2在Vicon场地使用同一检查点，四种偏好乘三种指令条件各五次，策略50Hz、底层200Hz。每段试验偏好固定，只有重放条件严格共享指令轨迹。", "resultsZh": "作者报告100偏好中67个采样行为非支配，平均偏好—目标相关0.843；实机不同偏好相对均衡设置最多降低单位距离能耗30.4%、位置误差38.7%、姿态峰值59%。这些最大值来自不同条件，不是一次运行同时实现。", "limitationsZh": "跟踪位置更好可伴随速度RMSE或能耗更差；慢速稳定偏好的速度误差为2.203米/秒。五次/条件不能支持密集真机可控性或统计显著性，非支配样本不是全球Pareto最优证明，手动扰动恢复仅定性。", "reproductionZh": "复现依赖完整奖励分解、偏好采样、固定先验和域随机化范围；附录提供网络、十步历史及控制配置。应分别核对实机四锚点与仿真100偏好，不能把接口可调直接表述为已验证连续在线切换。", "experimentType": "both", "robots": [ "Unitree Go2" ], "evidenceNotes": [ { "section": "III-A", "note": "所有报告评测每回合偏好固定，非支配集合为有限采样近似。" }, { "section": "V-A; Table III", "note": "四偏好×三条件×五试验；实机Go2。" }, { "section": "V-B; VI-C", "note": "收益存在交叉目标代价；实机不主张统计显著性。" } ], "corrections": [ "补充已核实Unitree Go2；最大降幅属于不同条件和指标。" ], "sourceVersion": "2610.01260v1", "originalSha256": "91068a3909fad60606b289d8e3432d205e0dc13440c32010f4882c2f923c0666" }, "experimentNote": "IsaacLab中4096个Go2平地环境训练，随机化摩擦、惯性、质心和扰动。模拟密集测试100偏好；实机Go2在Vicon场地使用同一检查点，四种偏好乘三种指令条件各五次，策略50Hz、底层200Hz。每段试验偏好固定，只有重放条件严格共享指令轨迹。", "contribution": "PROMO把速度跟踪、身体稳定、能耗效率分成可调语义目标，将步态与接触正则保持为固定先验。偏好向量同时输入策略和历史编码，三头价值网络分别学习语义回报；特权编码在训练中对齐可部署历史表示，部署只保留本体观测、速度估计与单一策略。", "whyUseful": "复现依赖完整奖励分解、偏好采样、固定先验和域随机化范围；附录提供网络、十步历史及控制配置。应分别核对实机四锚点与仿真100偏好，不能把接口可调直接表述为已验证连续在线切换。", "analysisVerifiedAt": "2026-10-04T13:45:34.458232+00:00" }, { "id": "arxiv-2610.01288", "title": "Interaction-Stiffness-Guided Basis Allocation in Dynamic Movement Primitives for Efficient Skill Transfer", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01288v1", "titleZh": "用于高效技能迁移的动态运动基元：交互刚度引导的基函数分配", "abstractZh": "SC-DMP把肌电和人体姿态估计的交互刚度，与跨示范轨迹低方差结合为关键阶段先验；STR-Net仅按先验一致性和时序平滑细化，不依赖外部关键性真值。根据累积关键性重排基函数中心，再坐标搜索带宽，保留DMP结构。\nKUKA iiwa拖动示教，四通道表面肌电2000Hz，运动约100Hz；Z、M各12条示范、5至30基函数，另做插拔、两段浸入和曲面擦拭，每任务四条示范并与六种运动基元比较。", "summary": "SC-DMP把肌电和人体姿态估计的交互刚度，与跨示范轨迹低方差结合为关键阶段先验；STR-Net仅按先验一致性和时序平滑细化，不依赖外部关键性真值。根据累积关键性重排基函数中心，再坐标搜索带宽，保留DMP结构。", "experimentType": "real", "robots": [ "KUKA iiwa" ], "tags": [ "技能学习", "动态运动基元", "示范学习", "轨迹泛化" ], "limitations": "核心假设是关键动作表现为较高刚度和较一致轨迹，可能漏掉必须柔顺的关键阶段。主要证据为参考轨迹拟合误差，不能直接换算任务成功或接触安全；加权指标使用同一关键性，需独立几何/任务区域验证。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01288v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01288v1", "note": "III; IV-A：关键性先验、传感配置及有限局部搜索。" }, { "url": "https://arxiv.org/html/2610.01288v1", "note": "IV-B Table II; IV-C Table VII; V：主要误差收益、神经细化增量与适用边界。" }, { "url": "https://arxiv.org/pdf/2610.01288v1", "note": "III; IV-A：关键性先验、传感配置及有限局部搜索。" }, { "url": "https://arxiv.org/pdf/2610.01288v1", "note": "IV-B Table II; IV-C Table VII; V：主要误差收益、神经细化增量与适用边界。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "技能学习 / 动态运动基元", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "基础理论", "模仿学习" ], "robotFilters": [ "KUKA iiwa" ], "original": { "id": "arxiv-2610.01288", "originalSourceUrl": "https://arxiv.org/pdf/2610.01288v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01288v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01288v1", "sha256": "75d4cc45c9414abbba86645a3e157aea9c22e5c0275eb0b906ca20ff5f98b70b", "bytes": 2925520, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:05.553007+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01288v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01288v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01288v1", "pages": 12, "sourceIdentityStatus": "title_match", "sourceTitle": "Interaction-Stiffness-Guided Basis Allocation in Dynamic Movement Primitives for Efficient Skill Transfer", "sourceVersion": "2610.01288v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.467403+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01288v1" }, "originalAnalysis": { "id": "arxiv-2610.01288", "sourceUrl": "https://arxiv.org/pdf/2610.01288v1", "sectionsRead": [ "III-A–C", "IV-A–C", "V", "Tables II,VII" ], "methodsZh": "SC-DMP把肌电和人体姿态估计的交互刚度，与跨示范轨迹低方差结合为关键阶段先验；STR-Net仅按先验一致性和时序平滑细化，不依赖外部关键性真值。根据累积关键性重排基函数中心，再坐标搜索带宽，保留DMP结构。", "experimentsZh": "KUKA iiwa拖动示教，四通道表面肌电2000Hz，运动约100Hz；Z、M各12条示范、5至30基函数，另做插拔、两段浸入和曲面擦拭，每任务四条示范并与六种运动基元比较。", "resultsZh": "字母48种配置—指标比较中45项改善，平均RMSE降21.42%、几何关键区降32.50%；三实机任务相对普通DMP关键区RMSE降19.93%–23.86%。STR-Net增量小，平均全局误差再降1.38%，且任务2关键区略变差。", "limitationsZh": "核心假设是关键动作表现为较高刚度和较一致轨迹，可能漏掉必须柔顺的关键阶段。主要证据为参考轨迹拟合误差，不能直接换算任务成功或接触安全；加权指标使用同一关键性，需独立几何/任务区域验证。", "reproductionZh": "分离中心重分配、带宽搜索和STR-Net贡献；保持基函数数量、GMR参考及时间对齐一致，优先复现低容量与少示范设置。记录肌电处理和刚度标定；搜索只保证容差下坐标局部最优。", "experimentType": "real", "robots": [ "KUKA iiwa" ], "corrections": [ "神经细化器提升幅度远小于整体基函数重分配，不应把全部收益归于STR-Net。" ], "evidenceNotes": [ { "note": "关键性先验、传感配置及有限局部搜索。", "section": "III; IV-A" }, { "note": "主要误差收益、神经细化增量与适用边界。", "section": "IV-B Table II; IV-C Table VII; V" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2610.01288v1", "originalSha256": "75d4cc45c9414abbba86645a3e157aea9c22e5c0275eb0b906ca20ff5f98b70b" }, "experimentNote": "KUKA iiwa拖动示教，四通道表面肌电2000Hz，运动约100Hz；Z、M各12条示范、5至30基函数，另做插拔、两段浸入和曲面擦拭，每任务四条示范并与六种运动基元比较。", "contribution": "SC-DMP把肌电和人体姿态估计的交互刚度，与跨示范轨迹低方差结合为关键阶段先验；STR-Net仅按先验一致性和时序平滑细化，不依赖外部关键性真值。根据累积关键性重排基函数中心，再坐标搜索带宽，保留DMP结构。", "whyUseful": "分离中心重分配、带宽搜索和STR-Net贡献；保持基函数数量、GMR参考及时间对齐一致，优先复现低容量与少示范设置。记录肌电处理和刚度标定；搜索只保证容差下坐标局部最优。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2610.01301", "title": "Continual Learning for 6-DoF Grasp Synthesis via Experience and Demonstrations", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01301v1", "titleZh": "通过经验与示范实现六自由度抓取生成的持续学习", "abstractZh": "在几何采样抓取候选之外加入演示回忆，使用局部点云BPS和MLP学习32维归一化嵌入。近邻成败证据以距离权重更新Beta后验并给抓取排序；部署时追加记忆而不重训主网络，演示既新增候选也影响评分。\nPyBullet评测443个未见物体、十类别，每类100次适应及最多5次演示；顺序类别学习检查遗忘。实机Franka Panda配平行夹爪和D435i，54物体、六组、超过1500次抓取；每类20个相同评测场景，并允许最多三视角回退。", "summary": "在几何采样抓取候选之外加入演示回忆，使用局部点云BPS和MLP学习32维归一化嵌入。近邻成败证据以距离权重更新Beta后验并给抓取排序；部署时追加记忆而不重训主网络，演示既新增候选也影响评分。", "experimentType": "both", "robots": [ "Franka Panda" ], "tags": [ "六自由度抓取", "持续学习", "经验记忆", "示范学习" ], "limitations": "只看几何无法识别外形相似却需不同抓法的实例，也不能恢复缺失或失真的深度。场景是适度遮挡而非交锁堆叠；记忆扩大后演示召回时间增长，未系统验证传感噪声和接触动力学改变。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01301v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/pdf/2610.01301v1", "note": "§3.2：32维嵌入、近邻Beta后验与在线追加记忆。" }, { "url": "https://arxiv.org/pdf/2610.01301v1", "note": "§4.2 Table 4; §5：硬件、1500余次实机抓取、混合场景与钳子失败案例。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "六自由度抓取 / 持续学习", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "操作与抓取", "模仿学习" ], "robotFilters": [ "Franka Panda" ], "original": { "id": "arxiv-2610.01301", "originalSourceUrl": "https://arxiv.org/pdf/2610.01301v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01301v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01301v1", "sha256": "0f3e9555814bee2c4d9b49034f191237796c26b4225d720d535a1dd6631d27a9", "bytes": 10164685, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:05.708860+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01301v1", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01301v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01301v1", "pages": 21, "sourceIdentityStatus": "title_match", "sourceTitle": "Continual Learning for 6-DoF Grasp Synthesis via Experience and Demonstrations", "sourceVersion": "2610.01301v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.464050+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01301v1" }, "originalAnalysis": { "id": "arxiv-2610.01301", "sourceUrl": "https://arxiv.org/pdf/2610.01301v1", "sectionsRead": [ "§3", "§4.1–4.2", "Tables 3–4", "§5" ], "methodsZh": "在几何采样抓取候选之外加入演示回忆，使用局部点云BPS和MLP学习32维归一化嵌入。近邻成败证据以距离权重更新Beta后验并给抓取排序；部署时追加记忆而不重训主网络，演示既新增候选也影响评分。", "experimentsZh": "PyBullet评测443个未见物体、十类别，每类100次适应及最多5次演示；顺序类别学习检查遗忘。实机Franka Panda配平行夹爪和D435i，54物体、六组、超过1500次抓取；每类20个相同评测场景，并允许最多三视角回退。", "resultsZh": "作者报告仿真顺序适应平均成功率97.8%，最差遗忘2.4个百分点。实机适应后五组超过90%，混合场景抓取成功89.6%、清场100%；钳子仅由62.5%升到68.3%，说明相似几何但不同摩擦或重心仍会冲突。", "limitationsZh": "只看几何无法识别外形相似却需不同抓法的实例，也不能恢复缺失或失真的深度。场景是适度遮挡而非交锁堆叠；记忆扩大后演示召回时间增长，未系统验证传感噪声和接触动力学改变。", "reproductionZh": "应严格分开适应与评测物体场景，固定失败标签、演示预算及视角回退。需同时报告每次抓取成功和整场清空率，并保留冻结编码器与记忆扩增设置；未独立重现实验。", "experimentType": "both", "robots": [ "Franka Panda" ], "evidenceNotes": [ { "section/page": "§3.2", "note": "32维嵌入、近邻Beta后验与在线追加记忆。" }, { "section/page": "§4.2 Table 4; §5", "note": "硬件、1500余次实机抓取、混合场景与钳子失败案例。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "PyBullet评测443个未见物体、十类别，每类100次适应及最多5次演示；顺序类别学习检查遗忘。实机Franka Panda配平行夹爪和D435i，54物体、六组、超过1500次抓取；每类20个相同评测场景，并允许最多三视角回退。", "contribution": "在几何采样抓取候选之外加入演示回忆，使用局部点云BPS和MLP学习32维归一化嵌入。近邻成败证据以距离权重更新Beta后验并给抓取排序；部署时追加记忆而不重训主网络，演示既新增候选也影响评分。", "whyUseful": "应严格分开适应与评测物体场景，固定失败标签、演示预算及视角回退。需同时报告每次抓取成功和整场清空率，并保留冻结编码器与记忆扩增设置；未独立重现实验。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2610.01310", "title": "Ex vivo breach detection using electrical conductivity during robotic pedicle drilling in the spine", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01310v1", "titleZh": "利用电导信号检测机器人脊柱椎弓根钻孔中的穿破：离体实验研究", "abstractZh": "在Kurobot末端安装带双极电极的螺纹钻头，以25Hz读取局部导电信号。最初数毫米估计基线并限制阈值上下界，同时监测最近2毫米内的突变，两路任一触发即停止钻进；阈值由既往离体数据离线调定。\n外科医生手动建立入口并导引机械臂，随后自动钻削51枚新鲜猪腰椎，盐水环境模拟导电条件。以探针、视觉、CT/微CT和视频核验停止位置；通常10N、30rpm，具体力可按骨质调整。", "summary": "在Kurobot末端安装带双极电极的螺纹钻头，以25Hz读取局部导电信号。最初数毫米估计基线并限制阈值上下界，同时监测最近2毫米内的突变，两路任一触发即停止钻进；阈值由既往离体数据离线调定。", "experimentType": "real", "robots": [ "Kurobot（定制钻削平台）" ], "tags": [ "医疗机器人", "离体实验", "生物阻抗感知", "安全检测" ], "limitations": "不能写成患者试验或完全自主手术，也不能把A/B合计当零骨皮质改变。作者排除临床不可接受的轨迹偏离，仅分析朝脊髓管的计划轨迹；猪生长软骨、体液和骨质差异可能混淆信号。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01310v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01310v1", "note": "§2.3：离体猪椎、手动导引后自动钻削。" }, { "url": "https://arxiv.org/html/2610.01310v1", "note": "§3 / Fig.8：6 A /45 B；均值0.7毫米，并非全部零侵入。" }, { "url": "https://arxiv.org/html/2610.01310v1", "note": "§4：排除了不符合预定轨迹的不可接受偏离。" }, { "url": "https://arxiv.org/html/2610.01310v1", "note": "PDF p.12, Fig.8 (visual inspection)：直方图各区间2、4、8、32、5例，共51；负/零端6例与A类数量一致，不能称全无变形。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "医疗机器人 / 离体实验", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "安全与评估" ], "robotFilters": [], "original": { "id": "arxiv-2610.01310", "originalSourceUrl": "https://arxiv.org/pdf/2610.01310v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01310v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01310v1", "sha256": "95c943c6928655c9e1a70da8c4ce62ffb94c4142b307f18eeb92b9183f64b70a", "bytes": 5126420, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:57.052736+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01310v1", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01310v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01310v1", "pages": 15, "sourceIdentityStatus": "title_match", "sourceTitle": "Ex vivo breach detection using electrical conductivity during robotic pedicle drilling in the spine", "sourceVersion": "2610.01310v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.456365+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01310v1" }, "originalAnalysis": { "id": "arxiv-2610.01310", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090114+00:00", "sourceUrl": "https://arxiv.org/html/2610.01310v1", "sectionsRead": [ "§2.1–2.5", "Algorithm 1", "§3", "Figs.5–8", "§4–5" ], "methodsZh": "在Kurobot末端安装带双极电极的螺纹钻头，以25Hz读取局部导电信号。最初数毫米估计基线并限制阈值上下界，同时监测最近2毫米内的突变，两路任一触发即停止钻进；阈值由既往离体数据离线调定。", "experimentsZh": "外科医生手动建立入口并导引机械臂，随后自动钻削51枚新鲜猪腰椎，盐水环境模拟导电条件。以探针、视觉、CT/微CT和视频核验停止位置；通常10N、30rpm，具体力可按骨质调整。", "resultsZh": "作者报告51次均归为A/B级，其中A为6次、B为45次；停止变形长度均值0.7毫米、标准差0.45毫米、范围−0.9至1.4毫米。100%只适用于纳入分析的这组离体试验。", "limitationsZh": "不能写成患者试验或完全自主手术，也不能把A/B合计当零骨皮质改变。作者排除临床不可接受的轨迹偏离，仅分析朝脊髓管的计划轨迹；猪生长软骨、体液和骨质差异可能混淆信号。", "reproductionZh": "复现须区分离线阈值选择和在线评估，公开纳入/排除、轨迹、力速度和测量定义，并进行更系统的人体相关验证。文中平台名为Kurobot，不能据引言提到的商业导航机器人推断本实验型号。", "experimentType": "real", "robots": [ "Kurobot（定制钻削平台）" ], "evidenceNotes": [ { "section/page": "§2.3", "note": "离体猪椎、手动导引后自动钻削。" }, { "section/page": "§3 / Fig.8", "note": "6 A /45 B；均值0.7毫米，并非全部零侵入。" }, { "section/page": "§4", "note": "排除了不符合预定轨迹的不可接受偏离。" }, { "section/page": "PDF p.12, Fig.8 (visual inspection)", "note": "直方图各区间2、4、8、32、5例，共51；负/零端6例与A类数量一致，不能称全无变形。" } ], "corrections": { "clinicalScope": "51次离体猪腰椎研究；非患者临床试验、非全自主手术。" } }, "experimentNote": "外科医生手动建立入口并导引机械臂，随后自动钻削51枚新鲜猪腰椎，盐水环境模拟导电条件。以探针、视觉、CT/微CT和视频核验停止位置；通常10N、30rpm，具体力可按骨质调整。", "contribution": "在Kurobot末端安装带双极电极的螺纹钻头，以25Hz读取局部导电信号。最初数毫米估计基线并限制阈值上下界，同时监测最近2毫米内的突变，两路任一触发即停止钻进；阈值由既往离体数据离线调定。", "whyUseful": "复现须区分离线阈值选择和在线评估，公开纳入/排除、轨迹、力速度和测量定义，并进行更系统的人体相关验证。文中平台名为Kurobot，不能据引言提到的商业导航机器人推断本实验型号。", "analysisVerifiedAt": "2026-10-04T13:50:35.090114+00:00" }, { "id": "arxiv-2610.01351", "title": "Is Success All You Need? Investigating the Impact of Input Perturbations on VLA Behaviour in Tabletop Manipulation Tasks", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01351v1", "titleZh": "只看成功就够了吗？研究输入扰动对桌面操作任务中视觉—语言—动作模型行为的影响", "abstractZh": "这是一套行为鲁棒性评估方法，补充二元成功率。它记录成功轨迹的仿真时长、末端/关节路径、均值与高分位加加速度、夹爪累计运动；比较每任务均值相对变化再取任务中位数，用MAD看波动，Mann-Whitney检验配BH多重比较校正。\n完整评测π0.5和VLANeXt的四个LIBERO套件，OpenVLA-OFT只测Spatial；选用跨套件联合微调检查点。基线各任务50回合，七类扰动按全部对应LIBERO-Plus变体各两回合，物体扰动例外。行为统计仅纳入成功回合，导数用10Hz仿真频率。", "summary": "这是一套行为鲁棒性评估方法，补充二元成功率。它记录成功轨迹的仿真时长、末端/关节路径、均值与高分位加加速度、夹爪累计运动；比较每任务均值相对变化再取任务中位数，用MAD看波动，Mann-Whitney检验配BH…", "experimentType": "sim", "robots": [ "未核实" ], "tags": [ "视觉语言动作模型", "鲁棒性评估", "桌面操作", "行为分析" ], "limitations": "仅统计成功轨迹会产生选择效应：扰动可能先淘汰长而曲折的恢复行为，使剩余成功轨迹显得更稳定。高层指标不能直接诊断掉物或误抓，未研究失败轨迹，也没有实机验证；仿真时长不包含模型推理延迟。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01351v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01351v1", "note": "III-B–C：任务均值变化取中位数；MWU配BH校正；MAD描述性变化。" }, { "url": "https://arxiv.org/html/2610.01351v1", "note": "IV：50基线回合，扰动按变体两回合，OFT仅Spatial。" }, { "url": "https://arxiv.org/html/2610.01351v1", "note": "V; Table II; VI：24.1%与35.3%为成功轨迹jerk指标变化，非成功率增益。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "视觉语言动作模型 / 鲁棒性评估", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "视觉语言动作", "操作与抓取", "安全与评估" ], "robotFilters": [], "original": { "id": "arxiv-2610.01351", "originalSourceUrl": "https://arxiv.org/pdf/2610.01351v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01351v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01351v1", "sha256": "162190985a1d39616e85e08077d75c671a4c66350c7fd97e8f9a2ca7def13272", "bytes": 612365, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:56.910874+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01351v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01351v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01351v1", "pages": 9, "sourceIdentityStatus": "title_match", "sourceTitle": "Is Success All You Need? Investigating the Impact of Input Perturbations on VLA Behaviour in Tabletop Manipulation Tasks", "sourceVersion": "2610.01351v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.451850+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01351v1" }, "originalAnalysis": { "id": "arxiv-2610.01351", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458230+00:00", "sourceUrl": "https://arxiv.org/html/2610.01351v1", "sectionsRead": [ "III-A–C", "IV Experimental Setup", "V-A–B", "Tables II–III", "VI Limitations" ], "methodsZh": "这是一套行为鲁棒性评估方法，补充二元成功率。它记录成功轨迹的仿真时长、末端/关节路径、均值与高分位加加速度、夹爪累计运动；比较每任务均值相对变化再取任务中位数，用MAD看波动，Mann-Whitney检验配BH多重比较校正。", "experimentsZh": "完整评测π0.5和VLANeXt的四个LIBERO套件，OpenVLA-OFT只测Spatial；选用跨套件联合微调检查点。基线各任务50回合，七类扰动按全部对应LIBERO-Plus变体各两回合，物体扰动例外。行为统计仅纳入成功回合，导数用10Hz仿真频率。", "resultsZh": "作者发现即便任务成功，摄像机、机器人初态及传感扰动仍改变动作质量；Spatial相机扰动下π0.5与OpenVLA-OFT平均末端jerk的任务中位相对增幅为24.1%和35.3%，显著增加分别涉及7/10和10/10任务。", "limitationsZh": "仅统计成功轨迹会产生选择效应：扰动可能先淘汰长而曲折的恢复行为，使剩余成功轨迹显得更稳定。高层指标不能直接诊断掉物或误抓，未研究失败轨迹，也没有实机验证；仿真时长不包含模型推理延迟。", "reproductionZh": "论文给出vla-reliability与VLA-eval整合入口。复现应保存逐步轨迹及控制频率、扰动到原任务的映射、成功筛选和多重检验范围，报告样本数与每任务分布，避免只复算一个聚合百分比。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section": "III-B–C", "note": "任务均值变化取中位数；MWU配BH校正；MAD描述性变化。" }, { "section": "IV", "note": "50基线回合，扰动按变体两回合，OFT仅Spatial。" }, { "section": "V; Table II; VI", "note": "24.1%与35.3%为成功轨迹jerk指标变化，非成功率增益。" } ], "corrections": [ "仅有仿真评估，不据Franka夹爪举例反推实机硬件。" ], "sourceVersion": "2610.01351v1", "originalSha256": "162190985a1d39616e85e08077d75c671a4c66350c7fd97e8f9a2ca7def13272" }, "experimentNote": "完整评测π0.5和VLANeXt的四个LIBERO套件，OpenVLA-OFT只测Spatial；选用跨套件联合微调检查点。基线各任务50回合，七类扰动按全部对应LIBERO-Plus变体各两回合，物体扰动例外。行为统计仅纳入成功回合，导数用10Hz仿真频率。", "contribution": "这是一套行为鲁棒性评估方法，补充二元成功率。它记录成功轨迹的仿真时长、末端/关节路径、均值与高分位加加速度、夹爪累计运动；比较每任务均值相对变化再取任务中位数，用MAD看波动，Mann-Whitney检验配BH多重比较校正。", "whyUseful": "论文给出vla-reliability与VLA-eval整合入口。复现应保存逐步轨迹及控制频率、扰动到原任务的映射、成功筛选和多重检验范围，报告样本数与每任务分布，避免只复算一个聚合百分比。", "analysisVerifiedAt": "2026-10-04T13:45:34.458230+00:00" }, { "id": "arxiv-2610.01397", "title": "Continue, Abort, or Fall: Viability-Aware Policy Selection (VAPS) for Safe Humanoid Acrobatics", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01397v1", "titleZh": "继续、中止还是跌落：用于安全人形机器人特技动作的可行性感知策略选择VAPS", "abstractZh": "分别训练完成动作、放弃动作但双脚落地、保护性跌倒三个专家；从相同仿真状态做反事实滚动，学习与候选策略相关的有限窗口存活预测。每步优先保留最有任务价值且仍可行的策略，撤离途中也可再次升级到保护跌倒。\nMJLab中Unitree G1前空翻和LimX Oli侧空翻，50Hz、通常五种子；与单备用、单一端到端安全策略及含历史的蒸馏学生比较。真实Oli另测试手动中止19次、保护跌倒16次及完整选择器18次。", "summary": "分别训练完成动作、放弃动作但双脚落地、保护性跌倒三个专家；从相同仿真状态做反事实滚动，学习与候选策略相关的有限窗口存活预测。每步优先保留最有任务价值且仍可行的策略，撤离途中也可再次升级到保护跌倒。", "experimentType": "both", "robots": [ "Unitree G1（仿真）", "LimX Oli（仿真及实机）" ], "tags": [ "人形机器人", "安全控制", "强化学习", "动态运动" ], "limitations": "83.3%实机数字混合完成翻转、安全中止与保护跌倒三种目标，不是翻转成功率，也不是零伤害保证。实机调整了阈值以诱发切换，试验数小；预测依赖仿真失败状态和进入备用策略状态的覆盖。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01397v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01397v1", "note": "III-B–C：策略条件可行性及反事实数据。" }, { "url": "https://arxiv.org/html/2610.01397v1", "note": "IV-C; Table III：任务安全权衡、18次硬件计分含义。" }, { "url": "https://arxiv.org/pdf/2610.01397v1", "note": "III-B–C：策略条件可行性及反事实数据。" }, { "url": "https://arxiv.org/pdf/2610.01397v1", "note": "IV-C; Table III：任务安全权衡、18次硬件计分含义。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "人形机器人 / 安全控制", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "安全与评估", "人形机器人", "强化学习" ], "robotFilters": [ "Unitree G1" ], "original": { "id": "arxiv-2610.01397", "originalSourceUrl": "https://arxiv.org/pdf/2610.01397v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01397v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01397v1", "sha256": "85a0a26ac6bf5a2a1af44a4e07ee9e6c89c771186684cd587b19c50bde3834ea", "bytes": 3185906, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:51.365283+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01397v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01397v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01397v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "Continue, Abort, or Fall: Viability-Aware Policy Selection (VAPS) for Safe Humanoid Acrobatics", "sourceVersion": "2610.01397v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.449189+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01397v1" }, "originalAnalysis": { "id": "arxiv-2610.01397", "sourceUrl": "https://arxiv.org/pdf/2610.01397v1", "sectionsRead": [ "III-A–D", "IV-A–E", "V", "Tables I–III" ], "methodsZh": "分别训练完成动作、放弃动作但双脚落地、保护性跌倒三个专家；从相同仿真状态做反事实滚动，学习与候选策略相关的有限窗口存活预测。每步优先保留最有任务价值且仍可行的策略，撤离途中也可再次升级到保护跌倒。", "experimentsZh": "MJLab中Unitree G1前空翻和LimX Oli侧空翻，50Hz、通常五种子；与单备用、单一端到端安全策略及含历史的蒸馏学生比较。真实Oli另测试手动中止19次、保护跌倒16次及完整选择器18次。", "resultsZh": "G1完整层级的动作成功59.6%，对原始62.1%损失2.5个百分点，头部接触降至1.6%；执行决策约0.29毫秒。Oli手动中止17/19成功，保护跌倒12/16；完整系统15/18达到所选专家目标。", "limitationsZh": "83.3%实机数字混合完成翻转、安全中止与保护跌倒三种目标，不是翻转成功率，也不是零伤害保证。实机调整了阈值以诱发切换，试验数小；预测依赖仿真失败状态和进入备用策略状态的覆盖。", "reproductionZh": "需分开标注任务成功、站立、头手接触和冲击大小；固定同一报警规则比较备用结构，检查预测器校准与分布外误报。先在仿真验证状态恢复与反事实标签，不能据此自行开展危险真人附近特技。", "experimentType": "both", "robots": [ "Unitree G1（仿真）", "LimX Oli（仿真及实机）" ], "corrections": [ "83.3%是各选中专家自身目标的综合达成率，不是空翻成功率。" ], "evidenceNotes": [ { "note": "策略条件可行性及反事实数据。", "section": "III-B–C" }, { "note": "任务安全权衡、18次硬件计分含义。", "section": "IV-C; Table III" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2610.01397v1", "originalSha256": "85a0a26ac6bf5a2a1af44a4e07ee9e6c89c771186684cd587b19c50bde3834ea" }, "experimentNote": "MJLab中Unitree G1前空翻和LimX Oli侧空翻，50Hz、通常五种子；与单备用、单一端到端安全策略及含历史的蒸馏学生比较。真实Oli另测试手动中止19次、保护跌倒16次及完整选择器18次。", "contribution": "分别训练完成动作、放弃动作但双脚落地、保护性跌倒三个专家；从相同仿真状态做反事实滚动，学习与候选策略相关的有限窗口存活预测。每步优先保留最有任务价值且仍可行的策略，撤离途中也可再次升级到保护跌倒。", "whyUseful": "需分开标注任务成功、站立、头手接触和冲击大小；固定同一报警规则比较备用结构，检查预测器校准与分布外误报。先在仿真验证状态恢复与反事实标签，不能据此自行开展危险真人附近特技。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2610.01477", "title": "ALFRED: Requirement-driven development of an open-source mobile manipulator for long-term plant monitoring", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01477v1", "titleZh": "ALFRED：面向长期植物监测、由需求驱动开发的开源移动操作机器人", "abstractZh": "依据长期监测需求迭代四个模块化构型，以AgileX Hunter 2.0底盘、Unitree Z1机械臂及可换传感器组件构成ALFRED。设计同时考量维修、供电、激光视野和机械臂可达空间；用统一机器人碰撞模型比较构型，而非提出新的学习策略。\n森林年测在29天完成528次穿行，采集31.68小时、4.9TB数据；森林作为农田替代场景。ALFRED 2.0在人工植物行间测试自主采集。可达分析对每构型检查3110400个位姿及八个IK初值，并计算不同姿态的LiDAR遮挡。", "summary": "依据长期监测需求迭代四个模块化构型，以AgileX Hunter 2.0底盘、Unitree Z1机械臂及可换传感器组件构成ALFRED。设计同时考量维修、供电、激光视野和机械臂可达空间；用统一机器人碰撞模型比…", "experimentType": "both", "robots": [ "ALFRED", "ALFRED 2.0", "AgileX Hunter 2.0", "Unitree Z1" ], "tags": [ "移动操作", "农业机器人", "长期部署", "开源硬件" ], "limitations": "森林实验底盘由操作员转向，机械臂自动扫动；文中所有报告运动均预先规划，尚无在线下一最佳视点闭环。底盘不防雨雪、机械臂相对编码器初始化可能偏移；静态遮挡分析假定平地且忽略线缆。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01477v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/pdf/2610.01477v1", "note": "§3–4; PDF pp9–12：硬件、森林采集数量、操作员转向及离线机械臂规划。" }, { "url": "https://arxiv.org/pdf/2610.01477v1", "note": "§10.2 Table 3; §11：可用工作空间分母、IK预算及硬件限制。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "移动操作 / 农业机器人", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "操作与抓取" ], "robotFilters": [ "ALFRED", "ALFRED 2.0", "AgileX Hunter 2.0", "Unitree Z1" ], "original": { "id": "arxiv-2610.01477", "originalSourceUrl": "https://arxiv.org/pdf/2610.01477v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01477v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01477v1", "sha256": "d84ee99466a8257fe6fb3ed1cc4378bf936c0b001681b8ceedb717ffc50d998e", "bytes": 19956185, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:47.539541+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01477v1", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01477v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01477v1", "pages": 36, "sourceIdentityStatus": "title_match", "sourceTitle": "ALFRED: Requirement-driven development of an open-source mobile manipulator for long-term plant monitoring", "sourceVersion": "2610.01477v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.445901+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01477v1" }, "originalAnalysis": { "id": "arxiv-2610.01477", "sourceUrl": "https://arxiv.org/pdf/2610.01477v1", "sectionsRead": [ "§3–4", "§10.1–10.4", "§11–12", "Tables 2–3" ], "methodsZh": "依据长期监测需求迭代四个模块化构型，以AgileX Hunter 2.0底盘、Unitree Z1机械臂及可换传感器组件构成ALFRED。设计同时考量维修、供电、激光视野和机械臂可达空间；用统一机器人碰撞模型比较构型，而非提出新的学习策略。", "experimentsZh": "森林年测在29天完成528次穿行，采集31.68小时、4.9TB数据；森林作为农田替代场景。ALFRED 2.0在人工植物行间测试自主采集。可达分析对每构型检查3110400个位姿及八个IK初值，并计算不同姿态的LiDAR遮挡。", "resultsZh": "作者报告可用空间占可达位姿比例由34.0%提高到60.0%、66.1%，不是占全部采样空间比例。全年计划采集未漏期，但电池退化后更换电池和发电机才恢复续航；各次构型切换约六小时。", "limitationsZh": "森林实验底盘由操作员转向，机械臂自动扫动；文中所有报告运动均预先规划，尚无在线下一最佳视点闭环。底盘不防雨雪、机械臂相对编码器初始化可能偏移；静态遮挡分析假定平地且忽略线缆。", "reproductionZh": "论文提供CAD、URDF、Docker软件与物料表入口；复现应区分森林半自主采集、人工植物自主行间任务及模型分析，先处理传感器时间戳、驱动包与变换树一致性。本次未装配或运行系统。", "experimentType": "both", "robots": [ "ALFRED", "ALFRED 2.0", "AgileX Hunter 2.0", "Unitree Z1" ], "evidenceNotes": [ { "section/page": "§3–4; PDF pp9–12", "note": "硬件、森林采集数量、操作员转向及离线机械臂规划。" }, { "section/page": "§10.2 Table 3; §11", "note": "可用工作空间分母、IK预算及硬件限制。" } ], "corrections": [ { "field": "experimentNote", "value": "森林半自主实机＋人工植物行间实机＋模型可达/遮挡分析", "reason": "没有自然农田长期全自主监测或在线规划验证。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04" }, "experimentNote": "森林年测在29天完成528次穿行，采集31.68小时、4.9TB数据；森林作为农田替代场景。ALFRED 2.0在人工植物行间测试自主采集。可达分析对每构型检查3110400个位姿及八个IK初值，并计算不同姿态的LiDAR遮挡。", "contribution": "依据长期监测需求迭代四个模块化构型，以AgileX Hunter 2.0底盘、Unitree Z1机械臂及可换传感器组件构成ALFRED。设计同时考量维修、供电、激光视野和机械臂可达空间；用统一机器人碰撞模型比较构型，而非提出新的学习策略。", "whyUseful": "论文提供CAD、URDF、Docker软件与物料表入口；复现应区分森林半自主采集、人工植物自主行间任务及模型分析，先处理传感器时间戳、驱动包与变换树一致性。本次未装配或运行系统。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2610.01505", "title": "OpenSpace Lab Solution to the IROS 2026 Indoor Exploration Competition", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01505v1", "titleZh": "OpenSpace Lab参加IROS 2026室内探索竞赛的方案", "abstractZh": "单机器人将预训练LaMa/MapEx地图补全用于区域级目标排序，以有限宽度束搜索平衡信息增益、转移和返航代价，局部规划仍检查真实观测。多机器人利用共享地图、队友意图和历史轨迹抑制重复探索，并用通信中继交接和剩余预算决定返航。\n竞赛公开七地图允许反复调参，私有十地图仅一次提交；地图按大小设1000/1500/2000步预算，各取四起点。多机器人测2至5台智能体，预算500/1000/1500步，需进入与基站可通信区域才能提交地图。", "summary": "单机器人将预训练LaMa/MapEx地图补全用于区域级目标排序，以有限宽度束搜索平衡信息增益、转移和返航代价，局部规划仍检查真实观测。多机器人利用共享地图、队友意图和历史轨迹抑制重复探索，并用通信中继交接和剩余…", "experimentType": "sim", "robots": [ "未核实" ], "tags": [ "自主探索", "多机器人协作", "地图补全", "路径规划" ], "limitations": "短篇系统报告含较多按预算定制的启发式，未提供独立组件消融、随机种子方差或真实平台闭环测试。公开和私有地图差异不能直接解释为同分布性能下降；没有实体机器人型号证据。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01505v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01505v1", "note": "§3：七公开/十私有地图、四起点、2–5智能体与步数预算。" }, { "url": "https://arxiv.org/html/2610.01505v1", "note": "Tables 1–2：覆盖及排名来自作者竞赛报告。" }, { "url": "https://arxiv.org/html/2610.01505v1", "note": "§4：接收完整论文后拟开源；未给实机实验。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "自主探索 / 多机器人协作", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "多机器人", "导航与建图" ], "robotFilters": [], "original": { "id": "arxiv-2610.01505", "originalSourceUrl": "https://arxiv.org/pdf/2610.01505v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01505v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01505v1", "sha256": "1d62007ac08414a9e3948ca7e9f02a253c3423c1bb8f109db84c420348f0d268", "bytes": 550954, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:47.004170+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01505v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01505v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01505v1", "pages": 5, "sourceIdentityStatus": "title_match", "sourceTitle": "OpenSpace Lab Solution to the IROS 2026 Indoor Exploration Competition", "sourceVersion": "2610.01505v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.430756+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01505v1" }, "originalAnalysis": { "id": "arxiv-2610.01505", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090111+00:00", "sourceUrl": "https://arxiv.org/html/2610.01505v1", "sectionsRead": [ "§2.1–2.2", "§3", "Tables 1–2", "§4" ], "methodsZh": "单机器人将预训练LaMa/MapEx地图补全用于区域级目标排序，以有限宽度束搜索平衡信息增益、转移和返航代价，局部规划仍检查真实观测。多机器人利用共享地图、队友意图和历史轨迹抑制重复探索，并用通信中继交接和剩余预算决定返航。", "experimentsZh": "竞赛公开七地图允许反复调参，私有十地图仅一次提交；地图按大小设1000/1500/2000步预算，各取四起点。多机器人测2至5台智能体，预算500/1000/1500步，需进入与基站可通信区域才能提交地图。", "resultsZh": "作者报告公开单机器人覆盖61.04%、第一名；私有单/多机器人分别39.53%和39.91%、均第三。私有中等地图单机器人42.53%排名第一；这些数字是特定预算与上传规则下的竞赛成绩。", "limitationsZh": "短篇系统报告含较多按预算定制的启发式，未提供独立组件消融、随机种子方差或真实平台闭环测试。公开和私有地图差异不能直接解释为同分布性能下降；没有实体机器人型号证据。", "reproductionZh": "复现应锁定竞赛地图、通信/上传规则、预算与全部启发式阈值，记录探索覆盖和成功上传的差别。正文声明完整论文接收后再开源，现有链接不等于本轮已核验完整可运行实现。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "§3", "note": "七公开/十私有地图、四起点、2–5智能体与步数预算。" }, { "section/page": "Tables 1–2", "note": "覆盖及排名来自作者竞赛报告。" }, { "section/page": "§4", "note": "接收完整论文后拟开源；未给实机实验。" } ] }, "experimentNote": "竞赛公开七地图允许反复调参，私有十地图仅一次提交；地图按大小设1000/1500/2000步预算，各取四起点。多机器人测2至5台智能体，预算500/1000/1500步，需进入与基站可通信区域才能提交地图。", "contribution": "单机器人将预训练LaMa/MapEx地图补全用于区域级目标排序，以有限宽度束搜索平衡信息增益、转移和返航代价，局部规划仍检查真实观测。多机器人利用共享地图、队友意图和历史轨迹抑制重复探索，并用通信中继交接和剩余预算决定返航。", "whyUseful": "复现应锁定竞赛地图、通信/上传规则、预算与全部启发式阈值，记录探索覆盖和成功上传的差别。正文声明完整论文接收后再开源，现有链接不等于本轮已核验完整可运行实现。", "analysisVerifiedAt": "2026-10-04T13:50:35.090111+00:00" }, { "id": "arxiv-2610.01559", "title": "Completion Aware Guidance for World Action Models", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01559v1", "titleZh": "面向世界动作模型的任务完成感知引导", "abstractZh": "CAG研究世界动作模型生成看似合理却不完成关键阶段的问题，如一直持物而不释放。它不重新训练，而汇总视频/动作查询对语言词元的跨注意力，稀疏选中当前重要词元，按有界系数放大其键和值，让短时采样更容易出现完成转移。理想完成奖励仅用于动机推导，采样并没有可用的真实完成指示器。\n对DreamZero的三个官方DROID仿真场景各做50次，对Fast-WAM在RoboTwin2.0中原报告低于90%的九任务各做30次；两骨干所有任务统一引导强度0.15，仅改变推理引导。另比较1.6秒短片与五秒生成范围。", "summary": "CAG研究世界动作模型生成看似合理却不完成关键阶段的问题，如一直持物而不释放。它不重新训练，而汇总视频/动作查询对语言词元的跨注意力，稀疏选中当前重要词元，按有界系数放大其键和值，让短时采样更容易出现完成转移。…", "experimentType": "sim", "robots": [ "未核实" ], "tags": [ "世界模型", "动作生成", "任务规划", "推理时引导" ], "limitations": "所选RoboTwin子集刻意排除已饱和任务，不能推为全基准结果；开微波炉及放罐入篮任务反而下降。样本量小、未报告实机验证；注意力亲和度是启发式替代而非已证明的完成敏感度。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01559v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01559v1", "note": "3; Equations 2–8：完成指标不可在采样取得，实际由跨注意力稀疏门控替代。" }, { "url": "https://arxiv.org/html/2610.01559v1", "note": "4; Table 1：三DROID模拟场景各50次，九RoboTwin任务各30次。" }, { "url": "https://arxiv.org/html/2610.01559v1", "note": "4; Figure 2c：79%和40%描述失败案例内的归因比例。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "世界模型 / 动作生成", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "世界模型" ], "robotFilters": [], "original": { "id": "arxiv-2610.01559", "originalSourceUrl": "https://arxiv.org/pdf/2610.01559v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01559v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01559v1", "sha256": "5537afade60a1e45278ff5285e61a29ba65650a754e165be7d807fba52ed31c4", "bytes": 3776669, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:36.563001+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01559v1", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01559v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01559v1", "pages": 6, "sourceIdentityStatus": "title_match", "sourceTitle": "Completion Aware Guidance for World Action Models", "sourceVersion": "2610.01559v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.429183+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01559v1" }, "originalAnalysis": { "id": "arxiv-2610.01559", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458226+00:00", "sourceUrl": "https://arxiv.org/html/2610.01559v1", "sectionsRead": [ "3 Method", "4 Experiments", "Figure 2", "Table 1", "5 Limitations" ], "methodsZh": "CAG研究世界动作模型生成看似合理却不完成关键阶段的问题，如一直持物而不释放。它不重新训练，而汇总视频/动作查询对语言词元的跨注意力，稀疏选中当前重要词元，按有界系数放大其键和值，让短时采样更容易出现完成转移。理想完成奖励仅用于动机推导，采样并没有可用的真实完成指示器。", "experimentsZh": "对DreamZero的三个官方DROID仿真场景各做50次，对Fast-WAM在RoboTwin2.0中原报告低于90%的九任务各做30次；两骨干所有任务统一引导强度0.15，仅改变推理引导。另比较1.6秒短片与五秒生成范围。", "resultsZh": "作者报告九任务平均64.4%升至70.0%，DreamZero三任务平均约69%升至75%；失败案例中被归因为未完成想象的比例79%降至40%。该比例的分母是失败案例，不是全部回合，不能解读为总失败率下降39点。", "limitationsZh": "所选RoboTwin子集刻意排除已饱和任务，不能推为全基准结果；开微波炉及放罐入篮任务反而下降。样本量小、未报告实机验证；注意力亲和度是启发式替代而非已证明的完成敏感度。", "reproductionZh": "复现需锁定原模型权重、官方仿真场景、采样层及词元门控实现，保留固定引导强度和逐任务随机种子；还应记录失败归因规则和逐任务成功计数。本分析未运行模型。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section": "3; Equations 2–8", "note": "完成指标不可在采样取得，实际由跨注意力稀疏门控替代。" }, { "section": "4; Table 1", "note": "三DROID模拟场景各50次，九RoboTwin任务各30次。" }, { "section": "4; Figure 2c", "note": "79%和40%描述失败案例内的归因比例。" } ], "corrections": [ "实验范围明确为sim；79%→40%不是整体失败率。" ], "sourceVersion": "2610.01559v1", "originalSha256": "5537afade60a1e45278ff5285e61a29ba65650a754e165be7d807fba52ed31c4" }, "experimentNote": "对DreamZero的三个官方DROID仿真场景各做50次，对Fast-WAM在RoboTwin2.0中原报告低于90%的九任务各做30次；两骨干所有任务统一引导强度0.15，仅改变推理引导。另比较1.6秒短片与五秒生成范围。", "contribution": "CAG研究世界动作模型生成看似合理却不完成关键阶段的问题，如一直持物而不释放。它不重新训练，而汇总视频/动作查询对语言词元的跨注意力，稀疏选中当前重要词元，按有界系数放大其键和值，让短时采样更容易出现完成转移。理想完成奖励仅用于动机推导，采样并没有可用的真实完成指示器。", "whyUseful": "复现需锁定原模型权重、官方仿真场景、采样层及词元门控实现，保留固定引导强度和逐任务随机种子；还应记录失败归因规则和逐任务成功计数。本分析未运行模型。", "analysisVerifiedAt": "2026-10-04T13:45:34.458226+00:00" }, { "id": "arxiv-2610.01612", "title": "ReCo: Response-Consistent Locomotion with Policy-Aware MPC for Legged Manipulation", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01612v1", "titleZh": "ReCo：面向足式移动操作、结合策略感知模型预测控制的响应一致性运动控制", "abstractZh": "用PPO将腿部策略塑造成对速度、高度、俯仰命令具有一致瞬态和步态周期响应，再辨识策略闭环响应模型。MPC据此联合优化底盘命令与六关节臂速度，预测命令滞后和步态振荡，而不是假设底盘瞬时精确执行。\nUnitree Go2加ARX X5，在Isaac Gym训练、MuJoCo起伏地形测试50条轨迹，各含正常及80N持续0.1秒推扰，共100次；实机全栈运行于Jetson Orin NX，另用动捕测八条轨迹。", "summary": "用PPO将腿部策略塑造成对速度、高度、俯仰命令具有一致瞬态和步态周期响应，再辨识策略闭环响应模型。MPC据此联合优化底盘命令与六关节臂速度，预测命令滞后和步态振荡，而不是假设底盘瞬时精确执行。", "experimentType": "both", "robots": [ "Unitree Go2", "ARX X5" ], "tags": [ "足式机器人", "移动操作", "强化学习", "模型预测控制" ], "limitations": "辨识改善位置却增大姿态误差，高度重复性没有改善。部分基线额外接纯追踪器，因此比较不能单独归因规划；推扰仅一种安排。无动捕任务的误差在机载估计世界坐标系计算，不能证明无里程计漂移。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01612v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01612v1", "note": "III; IV-A Table II：闭环响应建模、100次统一测试和控制设置。" }, { "url": "https://arxiv.org/html/2610.01612v1", "note": "IV-C; IV-F; V：消融权衡、实机误差及里程计证据范围。" }, { "url": "https://arxiv.org/pdf/2610.01612v1", "note": "III; IV-A Table II：闭环响应建模、100次统一测试和控制设置。" }, { "url": "https://arxiv.org/pdf/2610.01612v1", "note": "IV-C; IV-F; V：消融权衡、实机误差及里程计证据范围。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "足式机器人 / 移动操作", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "运动控制", "操作与抓取", "强化学习" ], "robotFilters": [ "Unitree Go2", "ARX X5" ], "original": { "id": "arxiv-2610.01612", "originalSourceUrl": "https://arxiv.org/pdf/2610.01612v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01612v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01612v1", "sha256": "50e765a2e16b9f5bffddf0d11f8f4254d9237fa039b99ccaef4b20de0a2886fc", "bytes": 5996657, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:36.694107+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01612v1", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01612v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01612v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "ReCo: Response-Consistent Locomotion with Policy-Aware MPC for Legged Manipulation", "sourceVersion": "2610.01612v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.425421+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01612v1" }, "originalAnalysis": { "id": "arxiv-2610.01612", "sourceUrl": "https://arxiv.org/pdf/2610.01612v1", "sectionsRead": [ "III-A–D", "IV-A–F", "V", "Tables I–IV" ], "methodsZh": "用PPO将腿部策略塑造成对速度、高度、俯仰命令具有一致瞬态和步态周期响应，再辨识策略闭环响应模型。MPC据此联合优化底盘命令与六关节臂速度，预测命令滞后和步态振荡，而不是假设底盘瞬时精确执行。", "experimentsZh": "Unitree Go2加ARX X5，在Isaac Gym训练、MuJoCo起伏地形测试50条轨迹，各含正常及80N持续0.1秒推扰，共100次；实机全栈运行于Jetson Orin NX，另用动捕测八条轨迹。", "resultsZh": "仿真位置RMSE0.135米、方向23.8度，分别较最优基线低28.7%、27.4%；成功75/100、跌倒3次。动捕实机误差0.076米、9.93度。响应塑形使归一化预测MSE下降59.6%。", "limitationsZh": "辨识改善位置却增大姿态误差，高度重复性没有改善。部分基线额外接纯追踪器，因此比较不能单独归因规划；推扰仅一种安排。无动捕任务的误差在机载估计世界坐标系计算，不能证明无里程计漂移。", "reproductionZh": "复现2×2塑形/辨识消融，保持相同轨迹和时钟；同时报告RMSE、完成进度和跌倒，因为提前失败仅统计已执行段。MPC为100Hz、1秒窗口、67节点，基线移植和预算也应公开。", "experimentType": "both", "robots": [ "Unitree Go2", "ARX X5" ], "corrections": [ "型号明确为Go2–ARX X5；位置改进不意味着所有姿态/重复性指标同步改善。" ], "evidenceNotes": [ { "note": "闭环响应建模、100次统一测试和控制设置。", "section": "III; IV-A Table II" }, { "note": "消融权衡、实机误差及里程计证据范围。", "section": "IV-C; IV-F; V" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2610.01612v1", "originalSha256": "50e765a2e16b9f5bffddf0d11f8f4254d9237fa039b99ccaef4b20de0a2886fc" }, "experimentNote": "Unitree Go2加ARX X5，在Isaac Gym训练、MuJoCo起伏地形测试50条轨迹，各含正常及80N持续0.1秒推扰，共100次；实机全栈运行于Jetson Orin NX，另用动捕测八条轨迹。", "contribution": "用PPO将腿部策略塑造成对速度、高度、俯仰命令具有一致瞬态和步态周期响应，再辨识策略闭环响应模型。MPC据此联合优化底盘命令与六关节臂速度，预测命令滞后和步态振荡，而不是假设底盘瞬时精确执行。", "whyUseful": "复现2×2塑形/辨识消融，保持相同轨迹和时钟；同时报告RMSE、完成进度和跌倒，因为提前失败仅统计已执行段。MPC为100Hz、1秒窗口、67节点，基线移植和预算也应公开。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2610.01644", "title": "SonarVoxNet: Diver Detection in 3D Bounding Box using 3D Sonar", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01644v1", "titleZh": "SonarVoxNet：利用三维声呐检测潜水员的三维包围框", "abstractZh": "将三维声呐点云体素化，经稀疏卷积和中心检测头输出位置、尺寸及连续6D全旋转，而非只预测偏航。声呐专用增强增加稀疏前景覆盖；可选因果时序细化分别修正中心、旋转和尺寸，不提升低置信候选或删除原检测。\nDiver3D来自七次天然洞潜采集，含36场景、37010帧与34568个标注；按场景分为21/7/8个训练、验证、测试场景。对比五种点云检测器及骨干×检测头消融，以两种子均值评价3D AP、朝向误差和SDS。", "summary": "将三维声呐点云体素化，经稀疏卷积和中心检测头输出位置、尺寸及连续6D全旋转，而非只预测偏航。声呐专用增强增加稀疏前景覆盖；可选因果时序细化分别修正中心、旋转和尺寸，不提升低置信候选或删除原检测。", "experimentType": "data", "robots": [ "未核实" ], "tags": [ "水下机器人", "三维声呐", "潜水员检测", "三维目标检测" ], "limitations": "评测是单类真实声呐离线检测，不是自主潜水机器人跟随控制；未核实采集载体型号。数据地点有限，非刚性人体的全身框标签有歧义，粗高度体素和声学物理尚未充分建模。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01644v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/pdf/2610.01644v1", "note": "§IV-C, §V：按场景划分、两种子及验证选模。" }, { "url": "https://arxiv.org/pdf/2610.01644v1", "note": "Tables II–IV; §VI-F：AP增益、约48.5度旋转误差与标签/声学限制。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "水下机器人 / 三维声呐", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "其他机器人研究" ], "robotFilters": [], "original": { "id": "arxiv-2610.01644", "originalSourceUrl": "https://arxiv.org/pdf/2610.01644v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01644v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01644v1", "sha256": "222df7708016ff8dc1162b8afc89960b4e8221bd2712309037aefea8493228bd", "bytes": 2081277, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:31.794993+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01644v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01644v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01644v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "SonarVoxNet: Diver Detection in 3D Bounding Box using 3D Sonar", "sourceVersion": "2610.01644v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.420449+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01644v1" }, "originalAnalysis": { "id": "arxiv-2610.01644", "sourceUrl": "https://arxiv.org/pdf/2610.01644v1", "sectionsRead": [ "§III", "§IV–VI", "Tables II–IV", "§VI-F" ], "methodsZh": "将三维声呐点云体素化，经稀疏卷积和中心检测头输出位置、尺寸及连续6D全旋转，而非只预测偏航。声呐专用增强增加稀疏前景覆盖；可选因果时序细化分别修正中心、旋转和尺寸，不提升低置信候选或删除原检测。", "experimentsZh": "Diver3D来自七次天然洞潜采集，含36场景、37010帧与34568个标注；按场景分为21/7/8个训练、验证、测试场景。对比五种点云检测器及骨干×检测头消融，以两种子均值评价3D AP、朝向误差和SDS。", "resultsZh": "作者报告无增强AP@0.35为0.515，增强后0.703，加细化后0.774；严格IoU0.50对应0.233、0.359、0.467。全SO(3)比偏航头明显有效；无增强朝向误差仍约48.5度，不能把高AP理解为高精度关节姿态恢复。", "limitationsZh": "评测是单类真实声呐离线检测，不是自主潜水机器人跟随控制；未核实采集载体型号。数据地点有限，非刚性人体的全身框标签有歧义，粗高度体素和声学物理尚未充分建模。", "reproductionZh": "需采用固定场景划分防相邻帧泄漏，按验证AP选模型，并分列原检测、增强与时序模块结果；保持裁剪范围和坐标轴约定。原文细化只用当前及过去帧，未独立执行验证。", "experimentType": "data", "robots": [], "evidenceNotes": [ { "section/page": "§IV-C, §V", "note": "按场景划分、两种子及验证选模。" }, { "section/page": "Tables II–IV; §VI-F", "note": "AP增益、约48.5度旋转误差与标签/声学限制。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "Diver3D来自七次天然洞潜采集，含36场景、37010帧与34568个标注；按场景分为21/7/8个训练、验证、测试场景。对比五种点云检测器及骨干×检测头消融，以两种子均值评价3D AP、朝向误差和SDS。", "contribution": "将三维声呐点云体素化，经稀疏卷积和中心检测头输出位置、尺寸及连续6D全旋转，而非只预测偏航。声呐专用增强增加稀疏前景覆盖；可选因果时序细化分别修正中心、旋转和尺寸，不提升低置信候选或删除原检测。", "whyUseful": "需采用固定场景划分防相邻帧泄漏，按验证AP选模型，并分列原检测、增强与时序模块结果；保持裁剪范围和坐标轴约定。原文细化只用当前及过去帧，未独立执行验证。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2610.01682", "title": "Beyond Leaderboard Scores: A Deployment-Focused Protocol for Interpretable Tracking Evaluation in Pedestrian-Centric Environments", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01682v1", "titleZh": "超越排行榜分数：面向行人环境部署的可解释跟踪评测协议", "abstractZh": "采用相同检测输入固定事件集合，把跟踪拆成初始化延迟、漏检期间连续正确位置/身份、间断后身份恢复和近邻关联，辅以HOTA与负载相关时延。PedRefTrack将发布预测的寿命和隐藏身份保留分开，另设真值辅助诊断版本。\n在JRDB测试集27序列上比较六个开源跟踪器及两种参考变体，使用统一Person-MinkUNet检测、STEAM-LO世界坐标和修正后的3D-IoU程序；Jetson Orin 64GB最大功耗模式测纯CPU跟踪步骤，另施加合成漏检/框扰动。", "summary": "采用相同检测输入固定事件集合，把跟踪拆成初始化延迟、漏检期间连续正确位置/身份、间断后身份恢复和近邻关联，辅以HOTA与负载相关时延。PedRefTrack将发布预测的寿命和隐藏身份保留分开，另设真值辅助诊断版…", "experimentType": "data", "robots": [ "未核实" ], "tags": [ "多目标跟踪", "机器人导航", "部署评测", "嵌入式计算" ], "limitations": "速度不含检测器或导航栈；修正了官方IoU后不能直接与JRDB榜单比较。单一数据集、检测器工作点与HOTA调参限制外推；真值辅助不是可部署方法，身份恢复也不代表漏检期间有可用位置。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01682v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01682v1", "note": "§IV-A：27序列，同检测；修正IoU后结果不可直接比较官方榜单。" }, { "url": "https://arxiv.org/html/2610.01682v1", "note": "Fig.5 / §V-B：一秒连续位置与身份成功和间断后恢复是不同指标。" }, { "url": "https://arxiv.org/html/2610.01682v1", "note": "Table II / Fig.6：无GPU，纯跟踪步骤速度，排除检测推理。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "多目标跟踪 / 机器人导航", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "安全与评估", "导航与建图" ], "robotFilters": [], "original": { "id": "arxiv-2610.01682", "originalSourceUrl": "https://arxiv.org/pdf/2610.01682v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01682v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01682v1", "sha256": "ed572fc24af8c8fb22b182d7d80f4cf1f5fa5782fd40af3d276e0c975ad63a7c", "bytes": 790812, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:26.131768+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01682v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01682v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01682v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "Beyond Leaderboard Scores: A Deployment-Focused Protocol for Interpretable Tracking Evaluation in Pedestrian-Centric Environments", "sourceVersion": "2610.01682v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.418071+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01682v1" }, "originalAnalysis": { "id": "arxiv-2610.01682", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090109+00:00", "sourceUrl": "https://arxiv.org/html/2610.01682v1", "sectionsRead": [ "§III", "§IV", "§V", "Table II", "Figs.5–7", "§VI" ], "methodsZh": "采用相同检测输入固定事件集合，把跟踪拆成初始化延迟、漏检期间连续正确位置/身份、间断后身份恢复和近邻关联，辅以HOTA与负载相关时延。PedRefTrack将发布预测的寿命和隐藏身份保留分开，另设真值辅助诊断版本。", "experimentsZh": "在JRDB测试集27序列上比较六个开源跟踪器及两种参考变体，使用统一Person-MinkUNet检测、STEAM-LO世界坐标和修正后的3D-IoU程序；Jetson Orin 64GB最大功耗模式测纯CPU跟踪步骤，另施加合成漏检/框扰动。", "resultsZh": "作者报告非真值方法HOTA仅24.26–29.67%，但一秒漏检期间连续同ID且位置正确的比例最高45.4%，PedRefTrack37.0%。后者75–80检测框负载约15.2 FPS，说明平均精度不足以描述部署能力。", "limitationsZh": "速度不含检测器或导航栈；修正了官方IoU后不能直接与JRDB榜单比较。单一数据集、检测器工作点与HOTA调参限制外推；真值辅助不是可部署方法，身份恢复也不代表漏检期间有可用位置。", "reproductionZh": "论文提供评测仓库，复现关键是固定检测、阈值和事件分母，并按序列bootstrap而非把帧当独立样本。应报告CPU核数与时延分位数，进一步通过下游导航评测验证实际影响。", "experimentType": "data", "robots": [], "evidenceNotes": [ { "section/page": "§IV-A", "note": "27序列，同检测；修正IoU后结果不可直接比较官方榜单。" }, { "section/page": "Fig.5 / §V-B", "note": "一秒连续位置与身份成功和间断后恢复是不同指标。" }, { "section/page": "Table II / Fig.6", "note": "无GPU，纯跟踪步骤速度，排除检测推理。" } ] }, "experimentNote": "在JRDB测试集27序列上比较六个开源跟踪器及两种参考变体，使用统一Person-MinkUNet检测、STEAM-LO世界坐标和修正后的3D-IoU程序；Jetson Orin 64GB最大功耗模式测纯CPU跟踪步骤，另施加合成漏检/框扰动。", "contribution": "采用相同检测输入固定事件集合，把跟踪拆成初始化延迟、漏检期间连续正确位置/身份、间断后身份恢复和近邻关联，辅以HOTA与负载相关时延。PedRefTrack将发布预测的寿命和隐藏身份保留分开，另设真值辅助诊断版本。", "whyUseful": "论文提供评测仓库，复现关键是固定检测、阈值和事件分母，并按序列bootstrap而非把帧当独立样本。应报告CPU核数与时延分位数，进一步通过下游导航评测验证实际影响。", "analysisVerifiedAt": "2026-10-04T13:50:35.090109+00:00" }, { "id": "arxiv-2610.01698", "title": "ActiveWAM: Evidence-Aware Active Vision for World-Action Models", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01698v1", "titleZh": "ActiveWAM：面向世界动作模型的证据感知主动视觉", "abstractZh": "ActiveWAM将任务相关历史保留与可执行头部取景共同建模。训练时用冻结Wan先验变换已观测历史，约束关键特征和可见动态，原始与变换历史共享动作/未来目标；视角适配器记录相机姿态、时间及有效性，统一生成双臂和云台动作。部署只用原始观察，不做在线反演或候选排序。\nRoboTwin-AV扩展50项任务，每项100成功示范，三环境种子下各条件100测试回合；另测TAVIS和固定相机基准。AirbotPlay双臂搭配两自由度IQR云台，在三项厨房组合任务各20次，对比五方法，阶段切换预先规定。", "summary": "ActiveWAM将任务相关历史保留与可执行头部取景共同建模。训练时用冻结Wan先验变换已观测历史，约束关键特征和可见动态，原始与变换历史共享动作/未来目标；视角适配器记录相机姿态、时间及有效性，统一生成双臂和…", "experimentType": "both", "robots": [ "AirbotPlay（固定基座双臂）", "GR1T2（TAVIS仿真）", "Reachy2（TAVIS仿真）" ], "tags": [ "主动视觉", "世界动作模型", "双臂操作", "相机控制" ], "limitations": "真机30次失败包括12次操作错误、十次过早停止、五次目标超视野及三次硬件问题。短历史、固定焦距及人工规定阶段切换限制长程自治；示范收集使用特权掩码/深度，测试不提供。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01698v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01698v1", "note": "3.2–3.4：仅训练期历史反演；未来视频辅助训练，部署不依赖解码。" }, { "url": "https://arxiv.org/html/2610.01698v1", "note": "4.1; Tables 2,6：RoboTwin-AV 50任务、三种子；复合扰动对照和因子消融。" }, { "url": "https://arxiv.org/html/2610.01698v1", "note": "4.4–4.5; Table 8：实机AirbotPlay，三组合任务各20次，30/60完整成功。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "主动视觉 / 世界动作模型", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "世界模型", "操作与抓取" ], "robotFilters": [ "Reachy2" ], "original": { "id": "arxiv-2610.01698", "originalSourceUrl": "https://arxiv.org/pdf/2610.01698v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01698v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01698v1", "sha256": "07746732e5e30a92a3679d2ebde7382800ee493cb36bd9d43c066db598b980c8", "bytes": 5183115, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:22.017350+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01698v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01698v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01698v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "ActiveWAM: Evidence-Aware Active Vision for World-Action Models", "sourceVersion": "2610.01698v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.415343+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01698v1" }, "originalAnalysis": { "id": "arxiv-2610.01698", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458224+00:00", "sourceUrl": "https://arxiv.org/html/2610.01698v1", "sectionsRead": [ "3.1–3.4", "4.1–4.5", "Tables 1–8", "5 Conclusion" ], "methodsZh": "ActiveWAM将任务相关历史保留与可执行头部取景共同建模。训练时用冻结Wan先验变换已观测历史，约束关键特征和可见动态，原始与变换历史共享动作/未来目标；视角适配器记录相机姿态、时间及有效性，统一生成双臂和云台动作。部署只用原始观察，不做在线反演或候选排序。", "experimentsZh": "RoboTwin-AV扩展50项任务，每项100成功示范，三环境种子下各条件100测试回合；另测TAVIS和固定相机基准。AirbotPlay双臂搭配两自由度IQR云台，在三项厨房组合任务各20次，对比五方法，阶段切换预先规定。", "resultsZh": "作者报告RoboTwin-AV复合扰动53.3%，原始历史基线41.7%；二乘二消融显示历史反演与主动取景共同收益。真机完整成功30/60即50%，第一阶段66.7%；不能把阶段成功率写成整段完成率。", "limitationsZh": "真机30次失败包括12次操作错误、十次过早停止、五次目标超视野及三次硬件问题。短历史、固定焦距及人工规定阶段切换限制长程自治；示范收集使用特权掩码/深度，测试不提供。", "reproductionZh": "复现需保持相同传感/动作接口、原始与变换对的目标一致性、窗口前划分训练测试，以及头部轨迹和相机标定。4090单次动作推理约165毫秒，动作分块和插值频率不能误作每步新推理频率。", "experimentType": "both", "robots": [ "AirbotPlay（固定基座双臂）", "GR1T2（TAVIS仿真）", "Reachy2（TAVIS仿真）" ], "evidenceNotes": [ { "section": "3.2–3.4", "note": "仅训练期历史反演；未来视频辅助训练，部署不依赖解码。" }, { "section": "4.1; Tables 2,6", "note": "RoboTwin-AV 50任务、三种子；复合扰动对照和因子消融。" }, { "section": "4.4–4.5; Table 8", "note": "实机AirbotPlay，三组合任务各20次，30/60完整成功。" } ], "corrections": [ "补充实机型号AirbotPlay；GR1T2、Reachy2只列仿真证据。" ], "sourceVersion": "2610.01698v1", "originalSha256": "07746732e5e30a92a3679d2ebde7382800ee493cb36bd9d43c066db598b980c8" }, "experimentNote": "RoboTwin-AV扩展50项任务，每项100成功示范，三环境种子下各条件100测试回合；另测TAVIS和固定相机基准。AirbotPlay双臂搭配两自由度IQR云台，在三项厨房组合任务各20次，对比五方法，阶段切换预先规定。", "contribution": "ActiveWAM将任务相关历史保留与可执行头部取景共同建模。训练时用冻结Wan先验变换已观测历史，约束关键特征和可见动态，原始与变换历史共享动作/未来目标；视角适配器记录相机姿态、时间及有效性，统一生成双臂和云台动作。部署只用原始观察，不做在线反演或候选排序。", "whyUseful": "复现需保持相同传感/动作接口、原始与变换对的目标一致性、窗口前划分训练测试，以及头部轨迹和相机标定。4090单次动作推理约165毫秒，动作分块和插值频率不能误作每步新推理频率。", "analysisVerifiedAt": "2026-10-04T13:45:34.458224+00:00" }, { "id": "arxiv-2610.01726", "title": "Query-Conditioned Articulation Estimation from a Single Image", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01726v1", "titleZh": "基于查询点的单图像关节运动参数估计", "abstractZh": "QueryArt接收已标定RGB和二维查询点，以冻结DINOv3特征、多尺度金字塔与可变形注意力预测关节类型、方向和相对轴偏移。偏移按查询深度归一化，规避单图绝对尺度不可辨识；测得一点深度后再恢复米制轴位置。\n训练混合三套真实扫描与两套合成数据，在四套留出集及完全未用于训练的HOI!、Arti4D评估。Spot机械臂配手眼RGB-D和Jetson Thor，16个家具运动部件、五类视角共57个计分试验。", "summary": "QueryArt接收已标定RGB和二维查询点，以冻结DINOv3特征、多尺度金字塔与可变形注意力预测关节类型、方向和相对轴偏移。偏移按查询深度归一化，规避单图绝对尺度不可辨识；测得一点深度后再恢复米制轴位置。", "experimentType": "real", "robots": [ "Boston Dynamics Spot（带机械臂）" ], "tags": [ "关节物体操作", "三维感知", "单图像估计", "移动操作" ], "limitations": "实机抓取点人工指定，抓取失败被排除于分母；70.2%不能理解为端到端自主家具打开率。极端视角和遮挡会退化，米制执行仍依赖深度；几何误差只对正确分类且有效样本统计。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01726v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01726v1", "note": "III-A–D; IV-B：尺度归一化、模型和条件指标定义。" }, { "url": "https://arxiv.org/html/2610.01726v1", "note": "IV-D：Spot、57次、人工查询与抓取失败排除。" }, { "url": "https://arxiv.org/pdf/2610.01726v1", "note": "III-A–D; IV-B：尺度归一化、模型和条件指标定义。" }, { "url": "https://arxiv.org/pdf/2610.01726v1", "note": "IV-D：Spot、57次、人工查询与抓取失败排除。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "关节物体操作 / 三维感知", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "操作与抓取" ], "robotFilters": [ "Boston Dynamics Spot" ], "original": { "id": "arxiv-2610.01726", "originalSourceUrl": "https://arxiv.org/pdf/2610.01726v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01726v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.01726v1", "sha256": "277143ac5001b83db468384be375aeb31b1aaf2f1f78a04d2cb684d47b773eed", "bytes": 17514406, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:22.317227+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01726v1", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01726v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01726v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "Query-Conditioned Articulation Estimation from a Single Image", "sourceVersion": "2610.01726v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.410705+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01726v1" }, "originalAnalysis": { "id": "arxiv-2610.01726", "sourceUrl": "https://arxiv.org/pdf/2610.01726v1", "sectionsRead": [ "III-A–E", "IV-A–E", "V", "Tables I–III; Fig.4" ], "methodsZh": "QueryArt接收已标定RGB和二维查询点，以冻结DINOv3特征、多尺度金字塔与可变形注意力预测关节类型、方向和相对轴偏移。偏移按查询深度归一化，规避单图绝对尺度不可辨识；测得一点深度后再恢复米制轴位置。", "experimentsZh": "训练混合三套真实扫描与两套合成数据，在四套留出集及完全未用于训练的HOI!、Arti4D评估。Spot机械臂配手眼RGB-D和Jetson Thor，16个家具运动部件、五类视角共57个计分试验。", "resultsZh": "分布外总体几何成功率HOI!69.08%、Arti4D80.05%；实机40/57即70.2%，判定需抽屉开20厘米或转动45度。多数总体指标优于比较模型，但个别旋转轴误差仍由3DOI领先；消融显示方向与轴定位存在权衡。", "limitationsZh": "实机抓取点人工指定，抓取失败被排除于分母；70.2%不能理解为端到端自主家具打开率。极端视角和遮挡会退化，米制执行仍依赖深度；几何误差只对正确分类且有效样本统计。", "reproductionZh": "必须同时报告含分类错误的总体成功与条件几何误差；按关节实体宏平均防止多视角重复加权。先复现查询深度归一化及轴符号不变损失，再在包含抓取失败的完整流程重新测成功率。", "experimentType": "real", "robots": [ "Boston Dynamics Spot（带机械臂）" ], "corrections": [ "明确实机Spot；另有真实/合成数据评测，但没有独立仿真机器人闭环实验。", "70.2%排除抓取失败，且查询抓取点为人工给定。" ], "evidenceNotes": [ { "note": "尺度归一化、模型和条件指标定义。", "section": "III-A–D; IV-B" }, { "note": "Spot、57次、人工查询与抓取失败排除。", "section": "IV-D" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2610.01726v1", "originalSha256": "277143ac5001b83db468384be375aeb31b1aaf2f1f78a04d2cb684d47b773eed" }, "experimentNote": "训练混合三套真实扫描与两套合成数据，在四套留出集及完全未用于训练的HOI!、Arti4D评估。Spot机械臂配手眼RGB-D和Jetson Thor，16个家具运动部件、五类视角共57个计分试验。", "contribution": "QueryArt接收已标定RGB和二维查询点，以冻结DINOv3特征、多尺度金字塔与可变形注意力预测关节类型、方向和相对轴偏移。偏移按查询深度归一化，规避单图绝对尺度不可辨识；测得一点深度后再恢复米制轴位置。", "whyUseful": "必须同时报告含分类错误的总体成功与条件几何误差；按关节实体宏平均防止多视角重复加权。先复现查询深度归一化及轴符号不变损失，再在包含抓取失败的完整流程重新测成功率。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2610.01742", "title": "World Motion Models: Flexible Sequence Modeling of SE(3) Trajectories", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01742v1", "titleZh": "世界运动模型：三维刚体位姿轨迹的灵活序列建模", "abstractZh": "以带语义、局部几何描述的SE(3)位姿轨迹为共同表示，将机器人连杆、物体、人体和动作目标放入同一令牌网格。逐令牌独立噪声的流匹配及非连续上下文令牌，允许通过更换已知/未知掩模执行未来预测、策略、补帧和重定向。\n按基准分别训练检查点，覆盖Language Table五类推物、TraceGen、OMOMO、人形重定向和更多定性规划。推物每类50回合、200步；G1轨迹送入TWIST在物理仿真中跟踪，测试30Hz完整输入、3Hz关键帧和噪声输入。", "summary": "以带语义、局部几何描述的SE(3)位姿轨迹为共同表示，将机器人连杆、物体、人体和动作目标放入同一令牌网格。逐令牌独立噪声的流匹配及非连续上下文令牌，允许通过更换已知/未知掩模执行未来预测、策略、补帧和重定向。", "experimentType": "sim", "robots": [ "Unitree G1（仿真）", "Panda（仿真）", "UR3（仿真）", "UR5（仿真）", "UR10（仿真）", "WidowX（仿真）", "Kinova（仿真，型号未细分）", "iiwa14（仿真）", "iiwa7（仿真）" ], "tags": [ "运动生成", "三维表征", "流匹配", "跨机体泛化" ], "limitations": "统一的是表示与推断形式，尚非跨全部基准联合训练的一个模型；依赖几何/深度和状态估计，未见形态不能直接泛化。真实视频是动作输入，图7明确显示后续为仿真运行，不能作G1实机部署证据。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01742v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/pdf/2610.01742v1", "note": "§3 opening; §5：各基准独立检查点、几何输入和未见形态限制。" }, { "url": "https://arxiv.org/pdf/2610.01742v1", "note": "PDF p9 Figure 7, Table 5：真实单目视频后为Physics Sim rollout，86.90%是仿真跟踪。" }, { "url": "https://arxiv.org/pdf/2610.01742v1", "note": "PDF p9 Figure 8：附加机械臂规划示例的文字标签为panda、UR3、UR5、UR10、widowx、kinova、iiwa14、iiwa7，均为模拟/运动学示例。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "运动生成 / 三维表征", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "其他机器人研究" ], "robotFilters": [ "Unitree G1", "Franka Panda", "Universal Robots UR5", "WidowX" ], "original": { "id": "arxiv-2610.01742", "originalSourceUrl": "https://arxiv.org/pdf/2610.01742v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01742v1", "sha256": "b9c49936d9545f1f106c2b36b2d7e0190018a429fcf9e7c94d64a2a5de89e298", "bytes": 6670644, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:12.450379+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01742v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01742v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01742v1", "pages": 22, "sourceIdentityStatus": "title_match", "sourceTitle": "World Motion Models: Flexible Sequence Modeling of SE(3) Trajectories", "sourceVersion": "2610.01742v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.394223+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01742v1" }, "originalAnalysis": { "id": "arxiv-2610.01742", "sourceUrl": "https://arxiv.org/pdf/2610.01742v1", "sectionsRead": [ "§2", "§3.1–3.6", "§5", "PDF p9 Figure 7 and Table 5" ], "methodsZh": "以带语义、局部几何描述的SE(3)位姿轨迹为共同表示，将机器人连杆、物体、人体和动作目标放入同一令牌网格。逐令牌独立噪声的流匹配及非连续上下文令牌，允许通过更换已知/未知掩模执行未来预测、策略、补帧和重定向。", "experimentsZh": "按基准分别训练检查点，覆盖Language Table五类推物、TraceGen、OMOMO、人形重定向和更多定性规划。推物每类50回合、200步；G1轨迹送入TWIST在物理仿真中跟踪，测试30Hz完整输入、3Hz关键帧和噪声输入。", "resultsZh": "作者报告推物宏平均87.2%；MPC使两种绝对目标任务由90%/76%提高到100%/80%。G1干净输入仿真序列成功率86.90%，重定向吞吐86.97帧/秒。TraceGen误差降低，但400轨迹预测耗时4.18秒，高于对照2.66秒。", "limitationsZh": "统一的是表示与推断形式，尚非跨全部基准联合训练的一个模型；依赖几何/深度和状态估计，未见形态不能直接泛化。真实视频是动作输入，图7明确显示后续为仿真运行，不能作G1实机部署证据。", "reproductionZh": "需固定每基准数据划分、坐标与旋转尺度、掩模及低层跟踪器；区分模型生成误差和物理跟踪成功率。本次已视觉核对PDF图7、表5及仿真标注，未运行代码。", "experimentType": "sim", "robots": [ "Unitree G1（仿真）", "Panda（仿真）", "UR3（仿真）", "UR5（仿真）", "UR10（仿真）", "WidowX（仿真）", "Kinova（仿真，型号未细分）", "iiwa14（仿真）", "iiwa7（仿真）" ], "evidenceNotes": [ { "section/page": "§3 opening; §5", "note": "各基准独立检查点、几何输入和未见形态限制。" }, { "section/page": "PDF p9 Figure 7, Table 5", "note": "真实单目视频后为Physics Sim rollout，86.90%是仿真跟踪。" }, { "section/page": "PDF p9 Figure 8", "note": "附加机械臂规划示例的文字标签为panda、UR3、UR5、UR10、widowx、kinova、iiwa14、iiwa7，均为模拟/运动学示例。" } ], "corrections": [ { "field": "experimentType", "value": "sim", "reason": "机器人闭环量化为仿真，另含真实视频离线数据；未确认实机。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04" }, "experimentNote": "按基准分别训练检查点，覆盖Language Table五类推物、TraceGen、OMOMO、人形重定向和更多定性规划。推物每类50回合、200步；G1轨迹送入TWIST在物理仿真中跟踪，测试30Hz完整输入、3Hz关键帧和噪声输入。", "contribution": "以带语义、局部几何描述的SE(3)位姿轨迹为共同表示，将机器人连杆、物体、人体和动作目标放入同一令牌网格。逐令牌独立噪声的流匹配及非连续上下文令牌，允许通过更换已知/未知掩模执行未来预测、策略、补帧和重定向。", "whyUseful": "需固定每基准数据划分、坐标与旋转尺度、掩模及低层跟踪器；区分模型生成误差和物理跟踪成功率。本次已视觉核对PDF图7、表5及仿真标注，未运行代码。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2610.01744", "title": "3DROID: A Renderable 3D Gaussian Dataset with Measured Per-Scene Reliability", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01744v1", "titleZh": "3DROID：具有逐场景可靠性测量的可渲染三维高斯数据集", "abstractZh": "利用机器人正运动学/CAD表面与经典双目深度的一致性先筛选相机外参，再用位姿条件化的前馈高斯重建生成可渲染场景。将外参可靠性、重建误差和测量覆盖分开公布，避免把图像逼真度直接等同真实几何。\n从DROID成功回合中构建520场景组，225组具备所需记录；精化外参门控保留114场景。三外视角重建、第四视角留出测试，110场景可进行图像评估；另用四视角检验机器人表面和双目深度一致性。", "summary": "利用机器人正运动学/CAD表面与经典双目深度的一致性先筛选相机外参，再用位姿条件化的前馈高斯重建生成可渲染场景。将外参可靠性、重建误差和测量覆盖分开公布，避免把图像逼真度直接等同真实几何。", "experimentType": "data", "robots": [ "未核实" ], "tags": [ "三维高斯", "机器人数据集", "相机标定", "数据可靠性" ], "limitations": "仅静态单帧、排除腕部视角，几何检查只覆盖有机器人表面或可靠双目支持处。114个发布场景中四个内参为零，另有一例借用中位内参；不能称全场景几何真值，也没有下游策略提升实验。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01744v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01744v1", "note": "§4.1：发布114，图像有效110；特殊内参病例有标注。" }, { "url": "https://arxiv.org/html/2610.01744v1", "note": "Table 4：分离位姿条件和外参精化，两因素联合增益不能只归因其一。" }, { "url": "https://arxiv.org/html/2610.01744v1", "note": "§5：无下游操控提升声明，无全场景几何真值。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "三维高斯 / 机器人数据集", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "数据集与基准" ], "robotFilters": [], "original": { "id": "arxiv-2610.01744", "originalSourceUrl": "https://arxiv.org/pdf/2610.01744v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01744v1", "sha256": "492c6bd17f5829458b43b191e9e5c10e14f8ecc62026352a1e6258442426b510", "bytes": 1643526, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:12.420884+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01744v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01744v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01744v1", "pages": 12, "sourceIdentityStatus": "title_match", "sourceTitle": "3DROID: A Renderable 3D Gaussian Dataset with Measured Per-Scene Reliability", "sourceVersion": "2610.01744v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.388521+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01744v1" }, "originalAnalysis": { "id": "arxiv-2610.01744", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090108+00:00", "sourceUrl": "https://arxiv.org/html/2610.01744v1", "sectionsRead": [ "§3", "§4.1–4.2", "Tables 1–4", "§5" ], "methodsZh": "利用机器人正运动学/CAD表面与经典双目深度的一致性先筛选相机外参，再用位姿条件化的前馈高斯重建生成可渲染场景。将外参可靠性、重建误差和测量覆盖分开公布，避免把图像逼真度直接等同真实几何。", "experimentsZh": "从DROID成功回合中构建520场景组，225组具备所需记录；精化外参门控保留114场景。三外视角重建、第四视角留出测试，110场景可进行图像评估；另用四视角检验机器人表面和双目深度一致性。", "resultsZh": "作者报告ZipSplat基线PSNR12.75，联合精化外参与位姿条件后16.27；固定外参源的逐场景PSNR中位增益为0.73–2.73 dB。几何改善并不一致，注入不可靠外参不能保证厘米级真实性。", "limitationsZh": "仅静态单帧、排除腕部视角，几何检查只覆盖有机器人表面或可靠双目支持处。114个发布场景中四个内参为零，另有一例借用中位内参；不能称全场景几何真值，也没有下游策略提升实验。", "reproductionZh": "复现需关联原DROID视频/关节状态、PointWorld精化外参和发布高斯资产；保持坐标约定、门控样本阈值30与比例区间0.85–1.15。外参变化同时改变评估几何，统计对照应在同外参源内进行。", "experimentType": "data", "robots": [], "evidenceNotes": [ { "section/page": "§4.1", "note": "发布114，图像有效110；特殊内参病例有标注。" }, { "section/page": "Table 4", "note": "分离位姿条件和外参精化，两因素联合增益不能只归因其一。" }, { "section/page": "§5", "note": "无下游操控提升声明，无全场景几何真值。" } ] }, "experimentNote": "从DROID成功回合中构建520场景组，225组具备所需记录；精化外参门控保留114场景。三外视角重建、第四视角留出测试，110场景可进行图像评估；另用四视角检验机器人表面和双目深度一致性。", "contribution": "利用机器人正运动学/CAD表面与经典双目深度的一致性先筛选相机外参，再用位姿条件化的前馈高斯重建生成可渲染场景。将外参可靠性、重建误差和测量覆盖分开公布，避免把图像逼真度直接等同真实几何。", "whyUseful": "复现需关联原DROID视频/关节状态、PointWorld精化外参和发布高斯资产；保持坐标约定、门控样本阈值30与比例区间0.85–1.15。外参变化同时改变评估几何，统计对照应在同外参源内进行。", "analysisVerifiedAt": "2026-10-04T13:50:35.090108+00:00" }, { "id": "arxiv-2610.01794", "title": "Continuous Conditioning of VLAs with Augmenting EMG and Visual Task Descriptors", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01794v1", "titleZh": "利用补充肌电与视觉任务描述对视觉语言动作模型进行连续条件控制", "abstractZh": "研究把任务描述持续注入语言以外的输入。EC-VLA在SmolVLA本体状态后拼接八通道肌电包络，用腕屈伸及放松指导左右修正和选择；VA-VLA在π0.5图像上用颜色分割掩码标出目标及放置点，两者文字仅给通用操作指令。\n实机SO-101由三位参与者各提供105段示范并分别训练；每模型测20次无杂物和十次杂物试验。视觉提示实验使用RoboCasa移动底盘Franka Panda仿真，16项任务约1600段示范，每任务50次，另选八任务加入干扰物。", "summary": "研究把任务描述持续注入语言以外的输入。EC-VLA在SmolVLA本体状态后拼接八通道肌电包络，用腕屈伸及放松指导左右修正和选择；VA-VLA在π0.5图像上用颜色分割掩码标出目标及放置点，两者文字仅给通用操作…", "experimentType": "both", "robots": [ "SO-101", "Franka Panda（RoboCasa仿真，移动底盘）" ], "tags": [ "视觉语言动作模型", "肌电接口", "多模态条件", "人机交互" ], "limitations": "肌电部署仍需要对应用户持续指导，未检验陌生用户；仿真掩码来自模拟器，真实分割误差未纳入。模态变化同时改变了反馈更新频率，不能将增益全部归因于非语言形式；逐任务成功率并非全都提高。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01794v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01794v1", "note": "III-A; VI-B：三人各105演示；实机SO-101，8通道Myo EMG。" }, { "url": "https://arxiv.org/html/2610.01794v1", "note": "III-B：视觉掩码实验为RoboCasa仿真，不是真机Franka结果。" }, { "url": "https://arxiv.org/html/2610.01794v1", "note": "Tables II–IV：成功率与部分完成分数不同；杂物任务Counter-to-Sink成功率0.18降至0.16。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "视觉语言动作模型 / 肌电接口", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "视觉语言动作" ], "robotFilters": [ "SO-101", "Franka Panda" ], "original": { "id": "arxiv-2610.01794", "originalSourceUrl": "https://arxiv.org/pdf/2610.01794v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01794v1", "sha256": "7ef7e2724a418bd6304900be351dc2009b5255244c6fd2454ab0f207a8ddff79", "bytes": 1745822, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:12.400951+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01794v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01794v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01794v1", "pages": 7, "sourceIdentityStatus": "title_match", "sourceTitle": "Continuous Conditioning of VLAs with Augmenting EMG and Visual Task Descriptors", "sourceVersion": "2610.01794v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.386452+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01794v1" }, "originalAnalysis": { "id": "arxiv-2610.01794", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458222+00:00", "sourceUrl": "https://arxiv.org/html/2610.01794v1", "sectionsRead": [ "III Methods", "IV Results", "V Conclusion", "Appendix VI-A–E", "Tables I–IV" ], "methodsZh": "研究把任务描述持续注入语言以外的输入。EC-VLA在SmolVLA本体状态后拼接八通道肌电包络，用腕屈伸及放松指导左右修正和选择；VA-VLA在π0.5图像上用颜色分割掩码标出目标及放置点，两者文字仅给通用操作指令。", "experimentsZh": "实机SO-101由三位参与者各提供105段示范并分别训练；每模型测20次无杂物和十次杂物试验。视觉提示实验使用RoboCasa移动底盘Franka Panda仿真，16项任务约1600段示范，每任务50次，另选八任务加入干扰物。", "resultsZh": "作者报告实机普通条件两者成功率均80%，杂物条件从30%升到50%；仿真全部任务均值由26.6%到31.4%，杂物八任务由11.3%到19.8%。完成分数包含部分进度，不能与完整成功率混用。", "limitationsZh": "肌电部署仍需要对应用户持续指导，未检验陌生用户；仿真掩码来自模拟器，真实分割误差未纳入。模态变化同时改变了反馈更新频率，不能将增益全部归因于非语言形式；逐任务成功率并非全都提高。", "reproductionZh": "附录给出肌电整流及2Hz低通、SmolVLA冻结骨干训练40K步，以及π0.5全参70K步、双H200训练。复现应保留参与者划分、持续人类输入和精确部分评分，另验证真实视觉跟踪器。", "experimentType": "both", "robots": [ "SO-101", "Franka Panda（RoboCasa仿真，移动底盘）" ], "evidenceNotes": [ { "section": "III-A; VI-B", "note": "三人各105演示；实机SO-101，8通道Myo EMG。" }, { "section": "III-B", "note": "视觉掩码实验为RoboCasa仿真，不是真机Franka结果。" }, { "section": "Tables II–IV", "note": "成功率与部分完成分数不同；杂物任务Counter-to-Sink成功率0.18降至0.16。" } ], "corrections": [ "机器人应区分实机SO-101与仿真Franka Panda；不能写成两种实机平台。" ], "sourceVersion": "2610.01794v1", "originalSha256": "7ef7e2724a418bd6304900be351dc2009b5255244c6fd2454ab0f207a8ddff79" }, "experimentNote": "实机SO-101由三位参与者各提供105段示范并分别训练；每模型测20次无杂物和十次杂物试验。视觉提示实验使用RoboCasa移动底盘Franka Panda仿真，16项任务约1600段示范，每任务50次，另选八任务加入干扰物。", "contribution": "研究把任务描述持续注入语言以外的输入。EC-VLA在SmolVLA本体状态后拼接八通道肌电包络，用腕屈伸及放松指导左右修正和选择；VA-VLA在π0.5图像上用颜色分割掩码标出目标及放置点，两者文字仅给通用操作指令。", "whyUseful": "附录给出肌电整流及2Hz低通、SmolVLA冻结骨干训练40K步，以及π0.5全参70K步、双H200训练。复现应保留参与者划分、持续人类输入和精确部分评分，另验证真实视觉跟踪器。", "analysisVerifiedAt": "2026-10-04T13:45:34.458222+00:00" }, { "id": "arxiv-2610.01849", "title": "FlashDexRetarget: Accelerating Dexterous Manipulation Data Generation through Multi-Motion Retargeting", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01849v1", "titleZh": "FlashDexRetarget：通过多动作重定向加速灵巧操作数据生成", "abstractZh": "把多条人手—物体参考动作合并训练同一跟踪策略，以未来参考、物体点云和手物距离描述接触；左右手各用独立actor–critic但观察完整双手状态。FlashSAC复用不同参考的经验，并扩展至5000万回放、1024宽critic。\nTACO、OakInk2、HOT3D清洁轨迹构成50动作主集，含单物体和双物体各25条，在IsaacSim比较XHand、Sharpa Wave Hand重定向；再扩展200/500/1000动作。实机回放成功轨迹展示擦板、倒入锅和关盖。", "summary": "把多条人手—物体参考动作合并训练同一跟踪策略，以未来参考、物体点云和手物距离描述接触；左右手各用独立actor–critic但观察完整双手状态。FlashSAC复用不同参考的经验，并扩展至5000万回放、102…", "experimentType": "both", "robots": [ "XHand（仿真）", "Sharpa Wave Hand（仿真）", "实机型号未明确核实" ], "tags": [ "灵巧操作", "运动重定向", "强化学习", "数据生成" ], "limitations": "统计是训练过程中累计找到并存储的成功轨迹，不是最终策略随机闭环成功率；平均姿态误差只在成功样本上计算。需要干净参考、准确几何及仿真特权状态；实机是回放验证，未给大量重复成功率或清楚标注硬件型号。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01849v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01849v1", "note": "III-D; IV-A Table I：算法容量、平台、计算成本和三种成功指标。" }, { "url": "https://arxiv.org/html/2610.01849v1", "note": "IV-E; V：真实回放三任务，特权状态与清洁轨迹限制。" }, { "url": "https://arxiv.org/pdf/2610.01849v1", "note": "III-D; IV-A Table I：算法容量、平台、计算成本和三种成功指标。" }, { "url": "https://arxiv.org/pdf/2610.01849v1", "note": "IV-E; V：真实回放三任务，特权状态与清洁轨迹限制。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "灵巧操作 / 运动重定向", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "操作与抓取", "灵巧手", "强化学习" ], "robotFilters": [ "Sharpa Wave" ], "original": { "id": "arxiv-2610.01849", "originalSourceUrl": "https://arxiv.org/pdf/2610.01849v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01849v1", "sha256": "31732324fd23a050541c610126d28822b6806a940a32701f3792fb74a867c082", "bytes": 16279287, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:05.743499+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01849v1", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01849v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01849v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "FlashDexRetarget: Accelerating Dexterous Manipulation Data Generation through Multi-Motion Retargeting", "sourceVersion": "2610.01849v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.382704+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01849v1" }, "originalAnalysis": { "id": "arxiv-2610.01849", "sourceUrl": "https://arxiv.org/pdf/2610.01849v1", "sectionsRead": [ "III-A–D", "IV-A–E", "V", "Table I; Figs 3–6" ], "methodsZh": "把多条人手—物体参考动作合并训练同一跟踪策略，以未来参考、物体点云和手物距离描述接触；左右手各用独立actor–critic但观察完整双手状态。FlashSAC复用不同参考的经验，并扩展至5000万回放、1024宽critic。", "experimentsZh": "TACO、OakInk2、HOT3D清洁轨迹构成50动作主集，含单物体和双物体各25条，在IsaacSim比较XHand、Sharpa Wave Hand重定向；再扩展200/500/1000动作。实机回放成功轨迹展示擦板、倒入锅和关盖。", "resultsZh": "严格同时约束手与物体的指标下，XHand成功86%，DexMachina12%；Sharpa为70%，Do as I Do10%。XHand主实验耗29GPU小时，对比66至2847。未来参考、接触奖励和分手网络均有贡献，PPO在相同步数下几乎无成功。", "limitationsZh": "统计是训练过程中累计找到并存储的成功轨迹，不是最终策略随机闭环成功率；平均姿态误差只在成功样本上计算。需要干净参考、准确几何及仿真特权状态；实机是回放验证，未给大量重复成功率或清楚标注硬件型号。", "reproductionZh": "保留失败动作的计算成本，分别报告SPIDER、MT和物体-only阈值。优先复现50动作与3亿步预算，再测试未见参考；不能将重定向教师直接宣称为仅视觉部署策略。", "experimentType": "both", "robots": [ "XHand（仿真）", "Sharpa Wave Hand（仿真）", "实机型号未明确核实" ], "corrections": [ "实机证据是轨迹回放，不是使用真实观测的闭环学习策略部署。" ], "evidenceNotes": [ { "note": "算法容量、平台、计算成本和三种成功指标。", "section": "III-D; IV-A Table I" }, { "note": "真实回放三任务，特权状态与清洁轨迹限制。", "section": "IV-E; V" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2610.01849v1", "originalSha256": "31732324fd23a050541c610126d28822b6806a940a32701f3792fb74a867c082" }, "experimentNote": "TACO、OakInk2、HOT3D清洁轨迹构成50动作主集，含单物体和双物体各25条，在IsaacSim比较XHand、Sharpa Wave Hand重定向；再扩展200/500/1000动作。实机回放成功轨迹展示擦板、倒入锅和关盖。", "contribution": "把多条人手—物体参考动作合并训练同一跟踪策略，以未来参考、物体点云和手物距离描述接触；左右手各用独立actor–critic但观察完整双手状态。FlashSAC复用不同参考的经验，并扩展至5000万回放、1024宽critic。", "whyUseful": "保留失败动作的计算成本，分别报告SPIDER、MT和物体-only阈值。优先复现50动作与3亿步预算，再测试未见参考；不能将重定向教师直接宣称为仅视觉部署策略。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2610.01856", "title": "ChunkVLA-AM: Parallel Action Chunking for Vision-Language-Action Robot Control in Additive Manufacturing", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01856v1", "titleZh": "ChunkVLA-AM：用于增材制造机器人控制的视觉语言动作模型并行动作分块", "abstractZh": "在OpenVLA-OFT上以LoRA适配单目RGB、语言和7维笛卡尔增量动作，预测八步动作块；云端推理、机器人端反归一化、平滑、限幅与校验。真实试验每执行完一块才重新观测请求，不是推理和执行并发。\n采用FAIRINO FR3固定工作站，以红蓝物块A到B转移模拟打印后取件。比较零样本和微调单步/分块配置的离线轨迹误差，另执行42次闭环实机转移，并测试七种物理灯光及十条轨迹的亮度扫描。", "summary": "在OpenVLA-OFT上以LoRA适配单目RGB、语言和7维笛卡尔增量动作，预测八步动作块；云端推理、机器人端反归一化、平滑、限幅与校验。真实试验每执行完一块才重新观测请求，不是推理和执行并发。", "experimentType": "real", "robots": [ "FAIRINO FR3" ], "tags": [ "视觉语言动作模型", "工业机器人", "动作分块", "增材制造" ], "limitations": "轨迹误差不能替代任务成功率；对比配置同时改变适配方式与预测时域，无法把改善单独归因于分块。场景固定、对象有限，限幅也不构成接触安全保证；亮度95不是跨任务最优照明。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01856v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01856v1", "note": "§IV-D–E：八步动作后才再次请求；服务器延迟不含全流程。" }, { "url": "https://arxiv.org/html/2610.01856v1", "note": "§V-D–E; §VI：FAIRINO FR3、1.74毫米离线MAE、39/42实机成功及混杂因素。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "视觉语言动作模型 / 工业机器人", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "视觉语言动作" ], "robotFilters": [ "FAIRINO FR3" ], "original": { "id": "arxiv-2610.01856", "originalSourceUrl": "https://arxiv.org/pdf/2610.01856v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01856v1", "sha256": "ebdddf3a7aa1b797383a056d45bf58477bf7678802f1abb6577aee97ab8178f0", "bytes": 1273702, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:05.712247+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01856v1", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01856v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01856v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "ChunkVLA-AM: Parallel Action Chunking for Vision-Language-Action Robot Control in Additive Manufacturing", "sourceVersion": "2610.01856v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.369789+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01856v1" }, "originalAnalysis": { "id": "arxiv-2610.01856", "sourceUrl": "https://arxiv.org/html/2610.01856v1", "sectionsRead": [ "§IV", "§V-B–F", "Tables I–II", "§VI" ], "methodsZh": "在OpenVLA-OFT上以LoRA适配单目RGB、语言和7维笛卡尔增量动作，预测八步动作块；云端推理、机器人端反归一化、平滑、限幅与校验。真实试验每执行完一块才重新观测请求，不是推理和执行并发。", "experimentsZh": "采用FAIRINO FR3固定工作站，以红蓝物块A到B转移模拟打印后取件。比较零样本和微调单步/分块配置的离线轨迹误差，另执行42次闭环实机转移，并测试七种物理灯光及十条轨迹的亮度扫描。", "resultsZh": "作者报告适配八步配置三轴平均MAE为1.74毫米，X/Y/Z分别0.37/0.97/3.87毫米；闭环独立计数39/42成功，即92.9%，三次失败为末端放置倾倒。服务器0.06–0.08秒仅模型推理，不含网络、采集及机械执行。", "limitationsZh": "轨迹误差不能替代任务成功率；对比配置同时改变适配方式与预测时域，无法把改善单独归因于分块。场景固定、对象有限，限幅也不构成接触安全保证；亮度95不是跨任务最优照明。", "reproductionZh": "原文给出LoRA秩32、学习率5e-4、有效批量32、约30轮及动作裁剪尺度；应分别复现离线预测和真实闭环，记录总延迟、放置高度与照明。未运行训练。", "experimentType": "real", "robots": [ "FAIRINO FR3" ], "evidenceNotes": [ { "section/page": "§IV-D–E", "note": "八步动作后才再次请求；服务器延迟不含全流程。" }, { "section/page": "§V-D–E; §VI", "note": "FAIRINO FR3、1.74毫米离线MAE、39/42实机成功及混杂因素。" } ], "corrections": [ { "field": "robots", "value": [ "FAIRINO FR3" ], "reason": "正文明确厂商，不可混同Franka FR3。" }, { "field": "methodNote", "value": "报告的物理实验采用顺序观测—推理—执行", "reason": "主机并行推理另行测试，未用于实机结果。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04" }, "experimentNote": "采用FAIRINO FR3固定工作站，以红蓝物块A到B转移模拟打印后取件。比较零样本和微调单步/分块配置的离线轨迹误差，另执行42次闭环实机转移，并测试七种物理灯光及十条轨迹的亮度扫描。", "contribution": "在OpenVLA-OFT上以LoRA适配单目RGB、语言和7维笛卡尔增量动作，预测八步动作块；云端推理、机器人端反归一化、平滑、限幅与校验。真实试验每执行完一块才重新观测请求，不是推理和执行并发。", "whyUseful": "原文给出LoRA秩32、学习率5e-4、有效批量32、约30轮及动作裁剪尺度；应分别复现离线预测和真实闭环，记录总延迟、放置高度与照明。未运行训练。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2610.01863", "title": "LiteReality-Agent: An Agentic System for Interactable 3D Indoor Scene Reconstruction", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01863v1", "titleZh": "LiteReality-Agent：用于可交互三维室内场景重建的智能体系统", "abstractZh": "把重建表达为可执行Room.py，代理利用标定图像、深度、点云和RoomPlan布局，按墙面材质、固定装置、小物体、程序资产分阶段编辑。碰撞与支撑检查分离，最后导出具有关节、碰撞体和估计惯性的MuJoCo场景。\n十个室内场景、620采集视角，与LiteReality及两个通用编码代理的RGB-only/+Scan配置比较。所有输出用相同相机和共享灯光渲染；611帧有共同有效深度，另有13名受试者完成的感知评分。", "summary": "把重建表达为可执行Room.py，代理利用标定图像、深度、点云和RoomPlan布局，按墙面材质、固定装置、小物体、程序资产分阶段编辑。碰撞与支撑检查分离，最后导出具有关节、碰撞体和估计惯性的MuJoCo场景。", "experimentType": "data", "robots": [ "未核实" ], "tags": [ "室内重建", "数字孪生", "具身仿真", "代码智能体" ], "limitations": "评估使用采集视角而非留出新视角，灯光取自本方法且未独立标定。零交叉排除物体内部及家具与建筑交叉；近支撑不要求所有腿接触。物理参数没有实测真值，不能据可导出仿真推断动力学真实。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01863v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01863v1", "note": "§4.1 / Appendix C：620个采集视角；不是新视角泛化评测。" }, { "url": "https://arxiv.org/html/2610.01863v1", "note": "Table 3 / Appendix D：10.50厘米；近接触与交叉指标有明确计数边界。" }, { "url": "https://arxiv.org/html/2610.01863v1", "note": "Appendix F：物理属性基于推断，缺少实测动力学验证。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "室内重建 / 数字孪生", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "其他机器人研究" ], "robotFilters": [], "original": { "id": "arxiv-2610.01863", "originalSourceUrl": "https://arxiv.org/pdf/2610.01863v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01863v1", "sha256": "b2e47d97a85159221186b3af0980b5f180e961f2fa406d676da817ba452e289d", "bytes": 45541214, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:05.747558+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01863v1", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01863v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01863v1", "pages": 26, "sourceIdentityStatus": "title_match", "sourceTitle": "LiteReality-Agent: An Agentic System for Interactable 3D Indoor Scene Reconstruction", "sourceVersion": "2610.01863v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.365496+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01863v1" }, "originalAnalysis": { "id": "arxiv-2610.01863", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090106+00:00", "sourceUrl": "https://arxiv.org/html/2610.01863v1", "sectionsRead": [ "§3.3–3.4", "§4.1–4.3", "Tables 2–3", "Appendices C–F" ], "methodsZh": "把重建表达为可执行Room.py，代理利用标定图像、深度、点云和RoomPlan布局，按墙面材质、固定装置、小物体、程序资产分阶段编辑。碰撞与支撑检查分离，最后导出具有关节、碰撞体和估计惯性的MuJoCo场景。", "experimentsZh": "十个室内场景、620采集视角，与LiteReality及两个通用编码代理的RGB-only/+Scan配置比较。所有输出用相同相机和共享灯光渲染；611帧有共同有效深度，另有13名受试者完成的感知评分。", "resultsZh": "作者表2报告PSNR14.129、LPIPS0.5392；最强+Scan基线为12.967/0.6250。深度MAE10.50厘米，86件家具均有底部近接触，检测到的家具间表面交叉为零。", "limitationsZh": "评估使用采集视角而非留出新视角，灯光取自本方法且未独立标定。零交叉排除物体内部及家具与建筑交叉；近支撑不要求所有腿接触。物理参数没有实测真值，不能据可导出仿真推断动力学真实。", "reproductionZh": "复现应保存每次代理预算、全部采集帧、渲染配置和物体计数；逐帧平均会让视图多的场景占更大权重。需额外测量质量/摩擦并进行交互验证后才用于可信控制评测，本轮未核验完整软件运行。", "experimentType": "data", "robots": [], "evidenceNotes": [ { "section/page": "§4.1 / Appendix C", "note": "620个采集视角；不是新视角泛化评测。" }, { "section/page": "Table 3 / Appendix D", "note": "10.50厘米；近接触与交叉指标有明确计数边界。" }, { "section/page": "Appendix F", "note": "物理属性基于推断，缺少实测动力学验证。" } ] }, "experimentNote": "十个室内场景、620采集视角，与LiteReality及两个通用编码代理的RGB-only/+Scan配置比较。所有输出用相同相机和共享灯光渲染；611帧有共同有效深度，另有13名受试者完成的感知评分。", "contribution": "把重建表达为可执行Room.py，代理利用标定图像、深度、点云和RoomPlan布局，按墙面材质、固定装置、小物体、程序资产分阶段编辑。碰撞与支撑检查分离，最后导出具有关节、碰撞体和估计惯性的MuJoCo场景。", "whyUseful": "复现应保存每次代理预算、全部采集帧、渲染配置和物体计数；逐帧平均会让视图多的场景占更大权重。需额外测量质量/摩擦并进行交互验证后才用于可信控制评测，本轮未核验完整软件运行。", "analysisVerifiedAt": "2026-10-04T13:50:35.090106+00:00" }, { "id": "arxiv-2610.01906", "title": "Towards Physical Underwater Robotic Assistance for Scuba Diver Movement in Confined Spaces", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01906v1", "titleZh": "面向受限空间中水肺潜水员运动的物理水下机器人辅助", "abstractZh": "RADMCS用气瓶外挂圆筒和两台T200推进器产生转向力觉，不需把执行器缠在潜水员肢体上。侧视相机通过AprilGrid与PnP估计离墙距离，指数滤波、异常跳变拒绝、死区和饱和比例控制把距离误差转为左右推进器PWM。\n八名不同潜水经验的参与者分别参与泳池及开放水域测试，并非人人完成全部条件。阈值实验比较225/325毫米力臂和纵横推进布局；蒙眼泳池沿七米标记墙游动测试离墙提示，安全潜水员全程跟随并可干预。", "summary": "RADMCS用气瓶外挂圆筒和两台T200推进器产生转向力觉，不需把执行器缠在潜水员肢体上。侧视相机通过AprilGrid与PnP估计离墙距离，指数滤波、异常跳变拒绝、死区和饱和比例控制把距离误差转为左右推进器P…", "experimentType": "real", "robots": [ "RADMCS（自研气瓶外挂水下辅助机器人）" ], "tags": [ "水下机器人", "可穿戴机器人", "人机交互", "力觉反馈" ], "limitations": "人数小、条件不完整，墙面标记受水流变形、自动对焦及水下光学影响；没有证明无标记洞穴内安全导航。壳体厂商100米额定值不等于整机已经通过100米工作验证，不能把可感知提示等同于安全自主辅助。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01906v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01906v1", "note": "3：两台Blue Robotics T200为推进器部件；整机名为RADMCS。" }, { "url": "https://arxiv.org/html/2610.01906v1", "note": "5; Table 2：八名参与者分别参加测试；阈值研究配置不完整。" }, { "url": "https://arxiv.org/html/2610.01906v1", "note": "5; Figures 9–10：距离丢失后控制卡在高左转命令，原文明确描述失败。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "水下机器人 / 可穿戴机器人", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "其他机器人研究" ], "robotFilters": [], "original": { "id": "arxiv-2610.01906", "originalSourceUrl": "https://arxiv.org/pdf/2610.01906v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01906v1", "sha256": "3430d9e4f33e3e99a2323da515f4e02ea4888626730fef9fee836f2145679c52", "bytes": 44802818, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:00.558828+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01906v1", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01906v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01906v1", "pages": 23, "sourceIdentityStatus": "title_match", "sourceTitle": "Towards Physical Underwater Robotic Assistance for Scuba Diver Movement in Confined Spaces", "sourceVersion": "2610.01906v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.316708+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01906v1" }, "originalAnalysis": { "id": "arxiv-2610.01906", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458220+00:00", "sourceUrl": "https://arxiv.org/html/2610.01906v1", "sectionsRead": [ "3 System Design", "4 Visual Control Methodology", "5 Human Study Evaluations", "Table 2", "Figures 5–10" ], "methodsZh": "RADMCS用气瓶外挂圆筒和两台T200推进器产生转向力觉，不需把执行器缠在潜水员肢体上。侧视相机通过AprilGrid与PnP估计离墙距离，指数滤波、异常跳变拒绝、死区和饱和比例控制把距离误差转为左右推进器PWM。", "experimentsZh": "八名不同潜水经验的参与者分别参与泳池及开放水域测试，并非人人完成全部条件。阈值实验比较225/325毫米力臂和纵横推进布局；蒙眼泳池沿七米标记墙游动测试离墙提示，安全潜水员全程跟随并可干预。", "resultsZh": "作者报告所测四人及三配置的平均感知阈值约为最大推力10%，展示了正确转向响应案例。论文也保留失败轨迹：距离估计约12秒后丢失，控制量未更新，持续给出高强度左转指令。", "limitationsZh": "人数小、条件不完整，墙面标记受水流变形、自动对焦及水下光学影响；没有证明无标记洞穴内安全导航。壳体厂商100米额定值不等于整机已经通过100米工作验证，不能把可感知提示等同于安全自主辅助。", "reproductionZh": "复现重点是机械推进器布局、PnP标定和滤波/死区参数、逐参与者协议及数据失效处理。人体水下实验涉及专门审批和安全保障；论文的控制原型不能直接作为实际潜水安全系统。", "experimentType": "real", "robots": [ "RADMCS（自研气瓶外挂水下辅助机器人）" ], "evidenceNotes": [ { "section": "3", "note": "两台Blue Robotics T200为推进器部件；整机名为RADMCS。" }, { "section": "5; Table 2", "note": "八名参与者分别参加测试；阈值研究配置不完整。" }, { "section": "5; Figures 9–10", "note": "距离丢失后控制卡在高左转命令，原文明确描述失败。" } ], "corrections": [ "区分真实泳池实验中的模拟贴墙场景与计算机仿真；实验类型为real。" ], "sourceVersion": "2610.01906v1", "originalSha256": "3430d9e4f33e3e99a2323da515f4e02ea4888626730fef9fee836f2145679c52" }, "experimentNote": "八名不同潜水经验的参与者分别参与泳池及开放水域测试，并非人人完成全部条件。阈值实验比较225/325毫米力臂和纵横推进布局；蒙眼泳池沿七米标记墙游动测试离墙提示，安全潜水员全程跟随并可干预。", "contribution": "RADMCS用气瓶外挂圆筒和两台T200推进器产生转向力觉，不需把执行器缠在潜水员肢体上。侧视相机通过AprilGrid与PnP估计离墙距离，指数滤波、异常跳变拒绝、死区和饱和比例控制把距离误差转为左右推进器PWM。", "whyUseful": "复现重点是机械推进器布局、PnP标定和滤波/死区参数、逐参与者协议及数据失效处理。人体水下实验涉及专门审批和安全保障；论文的控制原型不能直接作为实际潜水安全系统。", "analysisVerifiedAt": "2026-10-04T13:45:34.458220+00:00" }, { "id": "arxiv-2610.01910", "title": "Robot Learning on Discrete Surfaces: Theory and Applications", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01910v1", "titleZh": "离散曲面上的机器人学习：理论与应用", "abstractZh": "在多面体网格上定义切锥、最直测地线指数映射、对数映射与平行移动，显式处理边和顶点。DMP把驱动项固定在同一切锥再输运；GP使用测地距离核并检测长度尺度可容许性；流匹配沿原生网格几何积分。\n比较MeshDMP的跨曲面八字轨迹、粗细网格GP回归，以及Bunny/Cow表面50万样本的生成分布；实机是Franka Research 3，在四个3D打印曲面绘图，并用D405重建曲面、海绵加混合力阻抗控制清除粉笔线。", "summary": "在多面体网格上定义切锥、最直测地线指数映射、对数映射与平行移动，显式处理边和顶点。DMP把驱动项固定在同一切锥再输运；GP使用测地距离核并检测长度尺度可容许性；流匹配沿原生网格几何积分。", "experimentType": "both", "robots": [ "Franka Research 3" ], "tags": [ "几何机器人学习", "曲面运动", "模仿学习", "流匹配" ], "limitations": "超过约一万面时测地预计算成为瓶颈；距离核不能对任意长度尺度保证正定。位置轨迹之外的姿态需要离线平滑，力轨迹学习及动态曲面尚未解决，擦除示范也不等同工业材料去除验证。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01910v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01910v1", "note": "IV; V-C Table IV：原生算子、核正定边界、生成评估。" }, { "url": "https://arxiv.org/html/2610.01910v1", "note": "VI-A–B; VII：FR3、5N绘图力、海绵实机执行及复杂度限制。" }, { "url": "https://arxiv.org/pdf/2610.01910v1", "note": "IV; V-C Table IV：原生算子、核正定边界、生成评估。" }, { "url": "https://arxiv.org/pdf/2610.01910v1", "note": "VI-A–B; VII：FR3、5N绘图力、海绵实机执行及复杂度限制。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "几何机器人学习 / 曲面运动", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "模仿学习" ], "robotFilters": [ "Franka Research 3" ], "original": { "id": "arxiv-2610.01910", "originalSourceUrl": "https://arxiv.org/pdf/2610.01910v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01910v1", "sha256": "033aa88c33f537f886e8def34961e834a1acd8196a95e09d0827f296add06970", "bytes": 10416851, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:00.386673+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01910v1", "license": "http://creativecommons.org/licenses/by-sa/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01910v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01910v1", "pages": 16, "sourceIdentityStatus": "title_match", "sourceTitle": "Robot Learning on Discrete Surfaces: Theory and Applications", "sourceVersion": "2610.01910v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.259371+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01910v1" }, "originalAnalysis": { "id": "arxiv-2610.01910", "sourceUrl": "https://arxiv.org/pdf/2610.01910v1", "sectionsRead": [ "III–IV", "V-A–C", "VI-A–B", "VII", "Table IV; Figs 8–14" ], "methodsZh": "在多面体网格上定义切锥、最直测地线指数映射、对数映射与平行移动，显式处理边和顶点。DMP把驱动项固定在同一切锥再输运；GP使用测地距离核并检测长度尺度可容许性；流匹配沿原生网格几何积分。", "experimentsZh": "比较MeshDMP的跨曲面八字轨迹、粗细网格GP回归，以及Bunny/Cow表面50万样本的生成分布；实机是Franka Research 3，在四个3D打印曲面绘图，并用D405重建曲面、海绵加混合力阻抗控制清除粉笔线。", "resultsZh": "长时间积分保留八字极限环，避免基线旋转漂移；六组生成测试NLL均更低，例如Bunny第50特征函数0.76对1.48/1.55。相同10万更新的典型训练时间约20小时对30小时；四万轨迹点通常5秒内生成。", "limitationsZh": "超过约一万面时测地预计算成为瓶颈；距离核不能对任意长度尺度保证正定。位置轨迹之外的姿态需要离线平滑，力轨迹学习及动态曲面尚未解决，擦除示范也不等同工业材料去除验证。", "reproductionZh": "先测试边顶点处映射一致性与长时间周期漂移，再复现核矩阵正定检查；固定网格密度和采样分布比较NLL、运行时间。实机须区分几何轨迹迁移与已有力控器的作用。", "experimentType": "both", "robots": [ "Franka Research 3" ], "corrections": [ "原文不仅规划抛光动作，还用海绵和力阻抗控制执行粉笔线擦除；不能扩张为工业抛光性能验证。" ], "evidenceNotes": [ { "note": "原生算子、核正定边界、生成评估。", "section": "IV; V-C Table IV" }, { "note": "FR3、5N绘图力、海绵实机执行及复杂度限制。", "section": "VI-A–B; VII" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2610.01910v1", "originalSha256": "033aa88c33f537f886e8def34961e834a1acd8196a95e09d0827f296add06970" }, "experimentNote": "比较MeshDMP的跨曲面八字轨迹、粗细网格GP回归，以及Bunny/Cow表面50万样本的生成分布；实机是Franka Research 3，在四个3D打印曲面绘图，并用D405重建曲面、海绵加混合力阻抗控制清除粉笔线。", "contribution": "在多面体网格上定义切锥、最直测地线指数映射、对数映射与平行移动，显式处理边和顶点。DMP把驱动项固定在同一切锥再输运；GP使用测地距离核并检测长度尺度可容许性；流匹配沿原生网格几何积分。", "whyUseful": "先测试边顶点处映射一致性与长时间周期漂移，再复现核矩阵正定检查；固定网格密度和采样分布比较NLL、运行时间。实机须区分几何轨迹迁移与已有力控器的作用。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2610.01943", "title": "TouchTherm: Building Multimodal Digital Twins of Objects for Tactile and Thermal Rendering", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01943v1", "titleZh": "TouchTherm：构建用于触觉与热觉渲染的物体多模态数字孪生", "abstractZh": "把扫描粗网格与多光照法线图配准，在接触点切空间积分成微高度，叠加接触掩模渲染GelSight纹理。多视角红外视频提供温度观测，用图拉普拉斯物理约束网络识别表面扩散和环境换热系数，运行时直接积分温度场。\n为20件真实物体构建资产，触觉逐物体三次接触；热参数用独立60秒录制拟合，在留出视频上预测30、45秒，各300个重叠窗口。下游以纯模拟触觉训练ResNet-18，六种子测真实图像，另有10人VR温感演示。", "summary": "把扫描粗网格与多光照法线图配准，在接触点切空间积分成微高度，叠加接触掩模渲染GelSight纹理。多视角红外视频提供温度观测，用图拉普拉斯物理约束网络识别表面扩散和环境换热系数，运行时直接积分温度场。", "experimentType": "data", "robots": [ "未核实" ], "tags": [ "数字孪生", "触觉感知", "热觉交互", "多模态重建" ], "limitations": "配准含人工标注，采集每物体每模态约半天，未完整扫描底面；接触评测允许有限二维对齐。重叠热窗口不等于300个独立对象，VR小样本与分类示例也不是机器人闭环操作验证。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01943v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01943v1", "note": "§III-C–D：切空间触觉微几何与物理约束热场建模。" }, { "url": "https://arxiv.org/html/2610.01943v1", "note": "Tables I–II; §IV-D：20物体协议、热场误差、真实触觉分类；没有机器人闭环任务。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "数字孪生 / 触觉感知", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "人形机器人" ], "robotFilters": [], "original": { "id": "arxiv-2610.01943", "originalSourceUrl": "https://arxiv.org/pdf/2610.01943v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01943v1", "sha256": "2c4e10375711555a7555c5cfcf256d8387c418d0ffe899756438e3dbdd868b7f", "bytes": 3739119, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:33:56.057086+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01943v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01943v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01943v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "TouchTherm: Building Multimodal Digital Twins of Objects for Tactile and Thermal Rendering", "sourceVersion": "2610.01943v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.250186+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01943v1" }, "originalAnalysis": { "id": "arxiv-2610.01943", "sourceUrl": "https://arxiv.org/html/2610.01943v1", "sectionsRead": [ "§III-C–D", "§IV-A–D", "Tables I–II" ], "methodsZh": "把扫描粗网格与多光照法线图配准，在接触点切空间积分成微高度，叠加接触掩模渲染GelSight纹理。多视角红外视频提供温度观测，用图拉普拉斯物理约束网络识别表面扩散和环境换热系数，运行时直接积分温度场。", "experimentsZh": "为20件真实物体构建资产，触觉逐物体三次接触；热参数用独立60秒录制拟合，在留出视频上预测30、45秒，各300个重叠窗口。下游以纯模拟触觉训练ResNet-18，六种子测真实图像，另有10人VR温感演示。", "resultsZh": "作者报告微几何消融的G-SSIM/HF-NCC为0.0701/0.0911，优于粗网格0.0480/0.0116；30、45秒温度MAE为0.465、0.592℃。触觉分类Top-1由20%升至34%，仍有较大真实域差距；环境换热项贡献大于扩散项。", "limitationsZh": "配准含人工标注，采集每物体每模态约半天，未完整扫描底面；接触评测允许有限二维对齐。重叠热窗口不等于300个独立对象，VR小样本与分类示例也不是机器人闭环操作验证。", "reproductionZh": "应复现相机标定、独立触觉校准接触、热训练/测试录制拆分及物体等权汇总；保留传感器级颜色校准和分类训练隔离。未运行作者模型。", "experimentType": "data", "robots": [], "evidenceNotes": [ { "section/page": "§III-C–D", "note": "切空间触觉微几何与物理约束热场建模。" }, { "section/page": "Tables I–II; §IV-D", "note": "20物体协议、热场误差、真实触觉分类；没有机器人闭环任务。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "为20件真实物体构建资产，触觉逐物体三次接触；热参数用独立60秒录制拟合，在留出视频上预测30、45秒，各300个重叠窗口。下游以纯模拟触觉训练ResNet-18，六种子测真实图像，另有10人VR温感演示。", "contribution": "把扫描粗网格与多光照法线图配准，在接触点切空间积分成微高度，叠加接触掩模渲染GelSight纹理。多视角红外视频提供温度观测，用图拉普拉斯物理约束网络识别表面扩散和环境换热系数，运行时直接积分温度场。", "whyUseful": "应复现相机标定、独立触觉校准接触、热训练/测试录制拆分及物体等权汇总；保留传感器级颜色校准和分类训练隔离。未运行作者模型。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2610.01959", "title": "Training-Free Diffusion Planning with Analytical Local Scores", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01959v1", "titleZh": "基于解析局部得分的免训练扩散规划", "abstractZh": "将轨迹目标分解为平滑、障碍物、智能体间距和速度四类因子，直接构造局部解析得分，以退火Langevin更新修正粗规划轨迹。方法不学习示范分布，时间局部性只是带噪得分的近似，并非远端路径点严格独立。\n在Basic、Dense、Room、Shelf四种地图各测6/12/18个智能体，每设置25实例；900秒内无碰撞才算成功。对照MMD、SMD、DGD及MPPI，额外测试初始化、噪声、SDF引导和300智能体规模。", "summary": "将轨迹目标分解为平滑、障碍物、智能体间距和速度四类因子，直接构造局部解析得分，以退火Langevin更新修正粗规划轨迹。方法不学习示范分布，时间局部性只是带噪得分的近似，并非远端路径点严格独立。", "experimentType": "sim", "robots": [ "未核实" ], "tags": [ "运动规划", "扩散模型", "免训练方法", "多机器人协作" ], "limitations": "这是二维集中式路径规划，假设完美执行，未验证实体机器人或复杂动力学。结构地图中直线/Brownian初始化仅0–17%成功，RRT/Voronoi可达99–100%，故效果不能全归于扩散过程。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01959v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.01959v1", "note": "Table 1 caption：MPPI为GPU，其余CPU；不得把全部速度对照称为同GPU。" }, { "url": "https://arxiv.org/html/2610.01959v1", "note": "§5.2：A100上的300智能体均时5.39秒；初始化对结果关键。" }, { "url": "https://arxiv.org/html/2610.01959v1", "note": "§6 / Reproducibility：二维、集中式、完美执行假设；代码接受后公开。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "运动规划 / 扩散模型", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "多机器人" ], "robotFilters": [], "original": { "id": "arxiv-2610.01959", "originalSourceUrl": "https://arxiv.org/pdf/2610.01959v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01959v1", "sha256": "4d06a9baf70c6ca2b252f810891fb628d20d68f82ddf54ce36a5aefefd8c7b07", "bytes": 6682937, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:32:59.179894+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01959v1", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01959v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01959v1", "pages": 27, "sourceIdentityStatus": "title_match", "sourceTitle": "Training-Free Diffusion Planning with Analytical Local Scores", "sourceVersion": "2610.01959v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.243730+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01959v1" }, "originalAnalysis": { "id": "arxiv-2610.01959", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090104+00:00", "sourceUrl": "https://arxiv.org/html/2610.01959v1", "sectionsRead": [ "§4", "§5.1–5.2", "Tables 1–2", "§6", "Reproducibility statement" ], "methodsZh": "将轨迹目标分解为平滑、障碍物、智能体间距和速度四类因子，直接构造局部解析得分，以退火Langevin更新修正粗规划轨迹。方法不学习示范分布，时间局部性只是带噪得分的近似，并非远端路径点严格独立。", "experimentsZh": "在Basic、Dense、Room、Shelf四种地图各测6/12/18个智能体，每设置25实例；900秒内无碰撞才算成功。对照MMD、SMD、DGD及MPPI，额外测试初始化、噪声、SDF引导和300智能体规模。", "resultsZh": "作者报告12设置中11个成功率100%，Dense-18为96%；平均0.59秒，DGD为88.23秒。表1除MPPI使用GPU外其余为CPU；大规模300智能体、100障碍物平均5.39秒则在A100上。", "limitationsZh": "这是二维集中式路径规划，假设完美执行，未验证实体机器人或复杂动力学。结构地图中直线/Brownian初始化仅0–17%成功，RRT/Voronoi可达99–100%，故效果不能全归于扩散过程。", "reproductionZh": "论文给出250步、噪声/温度退火和附录扫参；需同时计入粗规划初始化并报告硬件。作者写明代码可索取、接受后公开，不能标成已公开完整训练代码；本方法本身无需训练。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "Table 1 caption", "note": "MPPI为GPU，其余CPU；不得把全部速度对照称为同GPU。" }, { "section/page": "§5.2", "note": "A100上的300智能体均时5.39秒；初始化对结果关键。" }, { "section/page": "§6 / Reproducibility", "note": "二维、集中式、完美执行假设；代码接受后公开。" } ], "corrections": { "codeStatus": "unknown", "trainingNote": "免训练规划；正文声明代码按请求提供并计划接收后公开。", "codeStatusNote": "本版本正文声明代码可索取、接受后公开；未核实后续发布。" } }, "experimentNote": "在Basic、Dense、Room、Shelf四种地图各测6/12/18个智能体，每设置25实例；900秒内无碰撞才算成功。对照MMD、SMD、DGD及MPPI，额外测试初始化、噪声、SDF引导和300智能体规模。", "contribution": "将轨迹目标分解为平滑、障碍物、智能体间距和速度四类因子，直接构造局部解析得分，以退火Langevin更新修正粗规划轨迹。方法不学习示范分布，时间局部性只是带噪得分的近似，并非远端路径点严格独立。", "whyUseful": "论文给出250步、噪声/温度退火和附录扫参；需同时计入粗规划初始化并报告硬件。作者写明代码可索取、接受后公开，不能标成已公开完整训练代码；本方法本身无需训练。", "analysisVerifiedAt": "2026-10-04T13:50:35.090104+00:00" }, { "id": "arxiv-2610.01985", "title": "H-SPAR: Hydrodynamic-aware Simulation for Particle Transport and Autonomous Robots", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.01985v1", "titleZh": "H-SPAR：融合水动力影响的粒子输运与自主机器人仿真", "abstractZh": "H-SPAR用离线Firedrake浅水方程流场统一驱动无人艇阻力和拉格朗日粒子输运，ROS2负责规划与水动力模块、Gazebo执行艇体物理。双船体按相对流速计算二次阻力和偏航力矩，采样概率按粒子相对速度偏离泵设计流速的程度衰减。\n比较RRT*、VF-RRT*及SVF-RRT*在相同静态地图上的计划和DWA闭环执行，区分平均流场与时变流场。覆盖实验比较0、45、90度扫掠，并逐项改变均匀/空间流场、确定/概率捕获及艇体受流力；每十秒注入100粒子。", "summary": "H-SPAR用离线Firedrake浅水方程流场统一驱动无人艇阻力和拉格朗日粒子输运，ROS2负责规划与水动力模块、Gazebo执行艇体物理。双船体按相对流速计算二次阻力和偏航力矩，采样概率按粒子相对速度偏离泵…", "experimentType": "sim", "robots": [ "双船体USV仿真模型（未注明商业型号）" ], "tags": [ "水面机器人", "水动力仿真", "路径规划", "环境采样" ], "limitations": "只有仿真证据。深度平均流场不解析垂向分层或三维湍流，单向耦合忽略艇尾流和螺旋桨扰动，概率捕获不是泵内细粒流体模型；物理数据校准仍列未来工作，逆流代价也不能直接视为电能节省。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.01985v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/pdf/2610.01985v1", "note": "2.1; 2.3; 2.5：浅水、双船体、单向流场耦合，非在线全CFD。" }, { "url": "https://arxiv.org/pdf/2610.01985v1", "note": "Table 2, p13：规划与执行代价、路线偏差分别列出。" }, { "url": "https://arxiv.org/pdf/2610.01985v1", "note": "Table 3, p14; Conclusion p16：完整配置速率0.11/0.09/0.10；实测验证属于未来工作。" }, { "url": "https://arxiv.org/pdf/2610.01985v1", "note": "PDF p14, Table 3：已渲染核对概率采样的总数与单位时间速率，45度总数最多但速率较低。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "水面机器人 / 水动力仿真", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "导航与建图" ], "robotFilters": [], "original": { "id": "arxiv-2610.01985", "originalSourceUrl": "https://arxiv.org/pdf/2610.01985v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.01985v1", "sha256": "909222aeff546168d34f2c0c8cd31cfdb1355d694b89c1195d220063b05080e3", "bytes": 4851330, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:32:54.755408+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.01985v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.01985v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.01985v1", "pages": 20, "sourceIdentityStatus": "title_match", "sourceTitle": "H-SPAR: Hydrodynamic-aware Simulation for Particle Transport and Autonomous Robots", "sourceVersion": "2610.01985v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.237418+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.01985v1" }, "originalAnalysis": { "id": "arxiv-2610.01985", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458227+00:00", "sourceUrl": "https://arxiv.org/pdf/2610.01985v1", "sectionsRead": [ "2.1–2.6", "3.1–3.2.1", "Tables 2–3", "4 Conclusion", "Data availability" ], "methodsZh": "H-SPAR用离线Firedrake浅水方程流场统一驱动无人艇阻力和拉格朗日粒子输运，ROS2负责规划与水动力模块、Gazebo执行艇体物理。双船体按相对流速计算二次阻力和偏航力矩，采样概率按粒子相对速度偏离泵设计流速的程度衰减。", "experimentsZh": "比较RRT*、VF-RRT*及SVF-RRT*在相同静态地图上的计划和DWA闭环执行，区分平均流场与时变流场。覆盖实验比较0、45、90度扫掠，并逐项改变均匀/空间流场、确定/概率捕获及艇体受流力；每十秒注入100粒子。", "resultsZh": "作者报告SVF-RRT*相对RRT*计划逆流代价降低69.4%，时变执行仅降低41.7%；表2中其执行路线更长且偏差6.7米，对照2.82米。完整采样配置的三方向速率为0.11、0.09、0.10，最佳总粒数与最佳单位时间速率并不相同。", "limitationsZh": "只有仿真证据。深度平均流场不解析垂向分层或三维湍流，单向耦合忽略艇尾流和螺旋桨扰动，概率捕获不是泵内细粒流体模型；物理数据校准仍列未来工作，逆流代价也不能直接视为电能节省。", "reproductionZh": "论文给出ROS2/Gazebo、DWA、流场生成及Zenodo数据入口。复现须固定流场时间插值、艇体阻力、粒子随机种子和采样半径，分别记录计划代价、真实执行轨迹及单位时间采样量；具体商业艇型未获全文支持。", "experimentType": "sim", "robots": [ "双船体USV仿真模型（未注明商业型号）" ], "evidenceNotes": [ { "section": "2.1; 2.3; 2.5", "note": "浅水、双船体、单向流场耦合，非在线全CFD。" }, { "section": "Table 2, p13", "note": "规划与执行代价、路线偏差分别列出。" }, { "section": "Table 3, p14; Conclusion p16", "note": "完整配置速率0.11/0.09/0.10；实测验证属于未来工作。" }, { "section": "PDF p14, Table 3", "note": "已渲染核对概率采样的总数与单位时间速率，45度总数最多但速率较低。" } ], "sourceVersion": "2610.01985v1", "originalSha256": "909222aeff546168d34f2c0c8cd31cfdb1355d694b89c1195d220063b05080e3" }, "experimentNote": "比较RRT*、VF-RRT*及SVF-RRT*在相同静态地图上的计划和DWA闭环执行，区分平均流场与时变流场。覆盖实验比较0、45、90度扫掠，并逐项改变均匀/空间流场、确定/概率捕获及艇体受流力；每十秒注入100粒子。", "contribution": "H-SPAR用离线Firedrake浅水方程流场统一驱动无人艇阻力和拉格朗日粒子输运，ROS2负责规划与水动力模块、Gazebo执行艇体物理。双船体按相对流速计算二次阻力和偏航力矩，采样概率按粒子相对速度偏离泵设计流速的程度衰减。", "whyUseful": "论文给出ROS2/Gazebo、DWA、流场生成及Zenodo数据入口。复现须固定流场时间插值、艇体阻力、粒子随机种子和采样半径，分别记录计划代价、真实执行轨迹及单位时间采样量；具体商业艇型未获全文支持。", "analysisVerifiedAt": "2026-10-04T13:45:34.458227+00:00" }, { "id": "arxiv-2610.02054", "title": "UniWAM: Unified World-Action Model", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.02054v1", "titleZh": "UniWAM：统一世界动作模型", "abstractZh": "以联合注意力连接Qwen3-VL物理推理、Wan2.2视觉生成及连续动作流匹配三个专家；人类动作被写成统一物理语言，机器人数据再提供精确动作监督。后训练扰动未来视觉潜变量，并从已执行动作历史加噪初始化动作生成。\n评测LIBERO、LIBERO-Plus、RoboTwin干净及随机化环境；实机用两台AgileX Piper，四项语言跟随任务及六阶段收桌。各实机任务200条示范，统一训练8万步、8张H100，评估每任务40次。", "summary": "以联合注意力连接Qwen3-VL物理推理、Wan2.2视觉生成及连续动作流匹配三个专家；人类动作被写成统一物理语言，机器人数据再提供精确动作监督。后训练扰动未来视觉潜变量，并从已执行动作历史加噪初始化动作生成。", "experimentType": "both", "robots": [ "AgileX Piper（双臂实机）" ], "tags": [ "世界动作模型", "视觉语言动作模型", "人机数据联合训练", "规模规律" ], "limitations": "RoboTwin干净条件并非所列方法最高；规模规律仅覆盖250至约5000小时机器人数据，少数据混入人类视频反会下降。t-SNE重叠是表示观察，不能单独证明因果迁移；实机任务规模有限。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.02054v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.02054v1", "note": "4; 5.2; Appendix B–D：模型组成、双Piper、40次评估及动作模板。" }, { "url": "https://arxiv.org/html/2610.02054v1", "note": "Tables 4–6; 5.2–5.4：量化成绩、规模区间、小规模负迁移和消融。" }, { "url": "https://arxiv.org/pdf/2610.02054v1", "note": "4; 5.2; Appendix B–D：模型组成、双Piper、40次评估及动作模板。" }, { "url": "https://arxiv.org/pdf/2610.02054v1", "note": "Tables 4–6; 5.2–5.4：量化成绩、规模区间、小规模负迁移和消融。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "世界动作模型 / 视觉语言动作模型", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "世界模型" ], "robotFilters": [ "AgileX PiPER" ], "original": { "id": "arxiv-2610.02054", "originalSourceUrl": "https://arxiv.org/pdf/2610.02054v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.02054v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.02054v1", "sha256": "d1c9eb01621c36f3f72a51f7cb1fb5cb457e7f24ba2b322ba4208097a09b1106", "bytes": 21229870, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:32:52.224224+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.02054v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.02054v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.02054v1", "pages": 29, "sourceIdentityStatus": "title_match", "sourceTitle": "UniWAM: Unified World-Action Model", "sourceVersion": "2610.02054v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.231578+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.02054v1" }, "originalAnalysis": { "id": "arxiv-2610.02054", "sourceUrl": "https://arxiv.org/pdf/2610.02054v1", "sectionsRead": [ "3", "4.1–4.3", "5.1–5.4", "Appendix B–D", "Tables 4–6" ], "methodsZh": "以联合注意力连接Qwen3-VL物理推理、Wan2.2视觉生成及连续动作流匹配三个专家；人类动作被写成统一物理语言，机器人数据再提供精确动作监督。后训练扰动未来视觉潜变量，并从已执行动作历史加噪初始化动作生成。", "experimentsZh": "评测LIBERO、LIBERO-Plus、RoboTwin干净及随机化环境；实机用两台AgileX Piper，四项语言跟随任务及六阶段收桌。各实机任务200条示范，统一训练8万步、8张H100，评估每任务40次。", "resultsZh": "LIBERO99.2%，RoboTwin干净75.14%、随机化68.32%；实机成功率67.5%、指令跟随率82.5%，分别高于π0.5约13.1、21.9个百分点。长任务进度5.0/6，仅略高于π0.5的4.8；历史初始化可用2–4步去噪。", "limitationsZh": "RoboTwin干净条件并非所列方法最高；规模规律仅覆盖250至约5000小时机器人数据，少数据混入人类视频反会下降。t-SNE重叠是表示观察，不能单独证明因果迁移；实机任务规模有限。", "reproductionZh": "先复现数据清洗、物理语言模板与动作历史对齐，再分离预训练来源、VLM解冻、未来噪声和历史初始化消融。分开记录指令理解、完整成功和阶段进度，避免把三种指标混用。", "experimentType": "both", "robots": [ "AgileX Piper（双臂实机）" ], "corrections": [ "实验类型由unknown修正为both；实机型号为AgileX Piper。" ], "evidenceNotes": [ { "note": "模型组成、双Piper、40次评估及动作模板。", "section": "4; 5.2; Appendix B–D" }, { "note": "量化成绩、规模区间、小规模负迁移和消融。", "section": "Tables 4–6; 5.2–5.4" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2610.02054v1", "originalSha256": "d1c9eb01621c36f3f72a51f7cb1fb5cb457e7f24ba2b322ba4208097a09b1106" }, "experimentNote": "评测LIBERO、LIBERO-Plus、RoboTwin干净及随机化环境；实机用两台AgileX Piper，四项语言跟随任务及六阶段收桌。各实机任务200条示范，统一训练8万步、8张H100，评估每任务40次。", "contribution": "以联合注意力连接Qwen3-VL物理推理、Wan2.2视觉生成及连续动作流匹配三个专家；人类动作被写成统一物理语言，机器人数据再提供精确动作监督。后训练扰动未来视觉潜变量，并从已执行动作历史加噪初始化动作生成。", "whyUseful": "先复现数据清洗、物理语言模板与动作历史对齐，再分离预训练来源、VLM解冻、未来噪声和历史初始化消融。分开记录指令理解、完整成功和阶段进度，避免把三种指标混用。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2610.02089", "title": "HumanoidToolBench: Benchmarking Humanoid Tool Use from Selection to Mobile Execution", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.02089v1", "titleZh": "HumanoidToolBench：从工具选择到移动执行的人形机器人工具使用基准", "abstractZh": "基准将工具使用拆为选择拾取、固定站位操作、移动操作三级，以推球、钩球及击碎冰块三种功能配合有无诱饵工具构成18任务。记录首次接触、正确工具抬升、工具接触目标及最终成功，避免只凭最终分数混淆选择与执行。\nToolBook含3003条仿真及91条真实演示。七种策略各训练一次，每个仿真任务100回合；三种策略经真实数据微调后，在G1与Dex3双手上测试推球、钩球四种L1条件，每项10次。实机未覆盖移动L2或碎冰全部任务。", "summary": "基准将工具使用拆为选择拾取、固定站位操作、移动操作三级，以推球、钩球及击碎冰块三种功能配合有无诱饵工具构成18任务。记录首次接触、正确工具抬升、工具接触目标及最终成功，避免只凭最终分数混淆选择与执行。", "experimentType": "both", "robots": [ "Unitree G1", "Dex3手" ], "tags": [ "人形机器人", "工具使用", "操作基准", "示范数据" ], "limitations": "真实样本小且场景范围有限，训练仅单次，不能据此给出稳定的全模型排名；不同策略采用各自动作块长度。L0、L1为不同任务，不能直接用两者成功率相减解释流程漏损。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.02089v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/pdf/2610.02089v1", "note": "§III-C–E; Table II：硬件、3094条数据及实机仅四种L1条件。" }, { "url": "https://arxiv.org/pdf/2610.02089v1", "note": "§IV-A; Tables III–IV：单次训练、每仿真任务100回合、真实每条件10次及结果。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "人形机器人 / 工具使用", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "操作与抓取", "数据集与基准", "人形机器人", "模仿学习" ], "robotFilters": [ "Unitree G1", "Dex3手" ], "original": { "id": "arxiv-2610.02089", "originalSourceUrl": "https://arxiv.org/pdf/2610.02089v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.02089v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.02089v1", "sha256": "d867b8a22da206cc4b9102c5246098fac73c9da6523d8ecd45eb34843b804693", "bytes": 10474149, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:32:47.688435+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.02089v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.02089v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.02089v1", "pages": 9, "sourceIdentityStatus": "title_match", "sourceTitle": "HumanoidToolBench: Benchmarking Humanoid Tool Use from Selection to Mobile Execution", "sourceVersion": "2610.02089v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.208308+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.02089v1" }, "originalAnalysis": { "id": "arxiv-2610.02089", "sourceUrl": "https://arxiv.org/pdf/2610.02089v1", "sectionsRead": [ "§III", "§IV-A–D", "Tables II–IV" ], "methodsZh": "基准将工具使用拆为选择拾取、固定站位操作、移动操作三级，以推球、钩球及击碎冰块三种功能配合有无诱饵工具构成18任务。记录首次接触、正确工具抬升、工具接触目标及最终成功，避免只凭最终分数混淆选择与执行。", "experimentsZh": "ToolBook含3003条仿真及91条真实演示。七种策略各训练一次，每个仿真任务100回合；三种策略经真实数据微调后，在G1与Dex3双手上测试推球、钩球四种L1条件，每项10次。实机未覆盖移动L2或碎冰全部任务。", "resultsZh": "作者报告FastWAM仿真标准推球由L1的76%降到L2的9%；GR00T N1.7真实标准推球6/10，FastWAM为5/10，ACT为1/10。诱饵通常降低强策略固定站位成绩，未见工具与无关指令探针揭示泛化和指令遵循问题。", "limitationsZh": "真实样本小且场景范围有限，训练仅单次，不能据此给出稳定的全模型排名；不同策略采用各自动作块长度。L0、L1为不同任务，不能直接用两者成功率相减解释流程漏损。", "reproductionZh": "需保留匹配初始场景、一分钟时限、失败与安全停止计数，并复用全身控制器。HTML在结果表截断，本次用原PDF补齐实验表与协议；未执行训练。", "experimentType": "both", "robots": [ "Unitree G1", "Dex3手" ], "evidenceNotes": [ { "section/page": "§III-C–E; Table II", "note": "硬件、3094条数据及实机仅四种L1条件。" }, { "section/page": "§IV-A; Tables III–IV", "note": "单次训练、每仿真任务100回合、真实每条件10次及结果。" } ], "corrections": [ { "field": "experimentNote", "value": "实机仅L1推球与钩球，两种工具模式；18任务全集为仿真", "reason": "§III-E与§IV-A限定真实覆盖。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04" }, "experimentNote": "ToolBook含3003条仿真及91条真实演示。七种策略各训练一次，每个仿真任务100回合；三种策略经真实数据微调后，在G1与Dex3双手上测试推球、钩球四种L1条件，每项10次。实机未覆盖移动L2或碎冰全部任务。", "contribution": "基准将工具使用拆为选择拾取、固定站位操作、移动操作三级，以推球、钩球及击碎冰块三种功能配合有无诱饵工具构成18任务。记录首次接触、正确工具抬升、工具接触目标及最终成功，避免只凭最终分数混淆选择与执行。", "whyUseful": "需保留匹配初始场景、一分钟时限、失败与安全停止计数，并复用全身控制器。HTML在结果表截断，本次用原PDF补齐实验表与协议；未执行训练。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2610.02110", "title": "GlassGuard: Verified Glass Plane Mapping for Robot Navigation", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.02110v1", "titleZh": "GlassGuard：面向机器人导航的经验证玻璃平面建图", "abstractZh": "先用裁剪量化的Slim SAM3生成玻璃掩码，再以LiDAR种子和水平/垂直结构构造有限平面。单帧射线几何、多视角重投影及地面证据排除假平面，最后合并全局假设并给规划地图补入占据体素。\n使用麦克纳姆轮平台、Livox Mid-360和Ricoh Theta相机，九场景覆盖六个室内外环境。超过2.1公里实机采集数据按原时间戳回放；统一输入比较MonoGlass3D、GlassRecon，并评估体素尺度、当前覆盖和误占据。", "summary": "先用裁剪量化的Slim SAM3生成玻璃掩码，再以LiDAR种子和水平/垂直结构构造有限平面。单帧射线几何、多视角重投影及地面证据排除假平面，最后合并全局假设并给规划地图补入占据体素。", "experimentType": "data", "robots": [ "麦克纳姆轮移动平台（型号未给出）" ], "tags": [ "机器人导航", "玻璃感知", "多传感器融合", "三维建图" ], "limitations": "Ever覆盖只表示曾检测，不能代替当前地图保护或碰撞率；0.5米体素下近距覆盖下降。主实验是录包回放，规划图是定性证据；仍依赖平面玻璃、标定和掩码质量，未证明闭环安全保证。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.02110v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.02110v1", "note": "§V-A：实机采集、随后各方法独立回放rosbags。" }, { "url": "https://arxiv.org/html/2610.02110v1", "note": "Tables III–IV：Ever/current coverage与体素尺度必须分别解释。" }, { "url": "https://arxiv.org/html/2610.02110v1", "note": "Fig.4：给出导航规划示例，而非实机碰撞率对照。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "机器人导航 / 玻璃感知", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "导航与建图" ], "robotFilters": [], "original": { "id": "arxiv-2610.02110", "originalSourceUrl": "https://arxiv.org/pdf/2610.02110v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.02110v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.02110v1", "sha256": "3c2c8f767dcf50db211313387be31c614832d420aada78bc37c217830a156f41", "bytes": 5019763, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:32:42.883143+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.02110v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.02110v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.02110v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "GlassGuard: Verified Glass Plane Mapping for Robot Navigation", "sourceVersion": "2610.02110v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.197103+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.02110v1" }, "originalAnalysis": { "id": "arxiv-2610.02110", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090100+00:00", "sourceUrl": "https://arxiv.org/html/2610.02110v1", "sectionsRead": [ "§III–IV", "§V-A–V-D", "Tables II–V", "Fig.4" ], "methodsZh": "先用裁剪量化的Slim SAM3生成玻璃掩码，再以LiDAR种子和水平/垂直结构构造有限平面。单帧射线几何、多视角重投影及地面证据排除假平面，最后合并全局假设并给规划地图补入占据体素。", "experimentsZh": "使用麦克纳姆轮平台、Livox Mid-360和Ricoh Theta相机，九场景覆盖六个室内外环境。超过2.1公里实机采集数据按原时间戳回放；统一输入比较MonoGlass3D、GlassRecon，并评估体素尺度、当前覆盖和误占据。", "resultsZh": "作者在1米体素下报告同针孔视野Ever覆盖82.1%，基线61.0%/44.4%；误体素每帧16.7，对比290.2/85.4。全景版本85.2%，RTX 5070 Ti笔记本GPU约0.73秒每帧。", "limitationsZh": "Ever覆盖只表示曾检测，不能代替当前地图保护或碰撞率；0.5米体素下近距覆盖下降。主实验是录包回放，规划图是定性证据；仍依赖平面玻璃、标定和掩码质量，未证明闭环安全保证。", "reproductionZh": "复现应保持相同5秒点云堆叠、视野、可见性筛选和人工真值；全景与针孔结果必须分开，消融的四场景标准差不能当置信区间。具体平台厂商型号未提供。", "experimentType": "data", "robots": [ "麦克纳姆轮移动平台（型号未给出）" ], "evidenceNotes": [ { "section/page": "§V-A", "note": "实机采集、随后各方法独立回放rosbags。" }, { "section/page": "Tables III–IV", "note": "Ever/current coverage与体素尺度必须分别解释。" }, { "section/page": "Fig.4", "note": "给出导航规划示例，而非实机碰撞率对照。" } ], "corrections": { "experimentType": "data", "hardwareNote": "有真实平台采集，但主定量证据为数据回放；不是闭环导航成功率实验。" } }, "experimentNote": "使用麦克纳姆轮平台、Livox Mid-360和Ricoh Theta相机，九场景覆盖六个室内外环境。超过2.1公里实机采集数据按原时间戳回放；统一输入比较MonoGlass3D、GlassRecon，并评估体素尺度、当前覆盖和误占据。", "contribution": "先用裁剪量化的Slim SAM3生成玻璃掩码，再以LiDAR种子和水平/垂直结构构造有限平面。单帧射线几何、多视角重投影及地面证据排除假平面，最后合并全局假设并给规划地图补入占据体素。", "whyUseful": "复现应保持相同5秒点云堆叠、视野、可见性筛选和人工真值；全景与针孔结果必须分开，消融的四场景标准差不能当置信区间。具体平台厂商型号未提供。", "analysisVerifiedAt": "2026-10-04T13:50:35.090100+00:00" }, { "id": "arxiv-2610.02120", "title": "SkeleWAM: Skeleton World-Action Modeling for Efficient Robotic Manipulation", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.02120v1", "titleZh": "SkeleWAM：面向高效机器人操作的骨架世界动作建模", "abstractZh": "冻结视觉网络从RGB-D提取物体中心及交互点，正运动学提供机器人节点，组成统一三维骨架。两个Transformer专家以流匹配联合学习动作和未来骨架，后者仅作训练监督。推理省去未来分支，MAC从多条采样轨迹中选距离其他轨迹最近的一条，避免直接平均不兼容动作。\n在含10,030变体、七类扰动的完整LIBERO-Plus测试零样本成功率。实机ARX R5配外置及腕部RealSense，相同五任务每方法各20次，对比Fast-WAM、π0.5、Cosmos-Policy；另做骨架、监督和推理消融。", "summary": "冻结视觉网络从RGB-D提取物体中心及交互点，正运动学提供机器人节点，组成统一三维骨架。两个Transformer专家以流匹配联合学习动作和未来骨架，后者仅作训练监督。推理省去未来分支，MAC从多条采样轨迹中选…", "experimentType": "both", "robots": [ "ARX R5" ], "tags": [ "世界动作模型", "三维表征", "机器人操作", "参数效率" ], "limitations": "布局扰动仅66.6%，低于π0.5的84.1%；真机样本有限，未证明复杂空间组合泛化。骨架依赖冻结感知和稳定节点对应，参数量优势不直接等于完整系统延迟优势。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.02120v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.02120v1", "note": "4.1; Table 1：LIBERO-Plus完整10,030变体；sim-state为特权输入。" }, { "url": "https://arxiv.org/html/2610.02120v1", "note": "4.3; Table 2; Figure 4：ARX R5真机五任务各20次，均值89%。" }, { "url": "https://arxiv.org/html/2610.02120v1", "note": "4.4; Tables 3–4：未来监督和MAC贡献分别通过80.1%、84.2%消融检验。" }, { "url": "https://arxiv.org/html/2610.02120v1", "note": "PDF p6, Figure 4; Table 1：已渲染核对ARX R5标识及观测输入/特权状态两行，未把相机作机器人型号。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "世界动作模型 / 三维表征", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "世界模型", "操作与抓取" ], "robotFilters": [ "ARX R5" ], "original": { "id": "arxiv-2610.02120", "originalSourceUrl": "https://arxiv.org/pdf/2610.02120v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.02120v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.02120v1", "sha256": "2518572d4baff871b3667fb5d22fd171f89ee6e87f192697a4e86f872adc4e49", "bytes": 3680379, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:32:39.475002+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.02120v1", "license": "http://creativecommons.org/licenses/by-nc-nd/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.02120v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.02120v1", "pages": 9, "sourceIdentityStatus": "title_match", "sourceTitle": "SkeleWAM: Skeleton World-Action Modeling for Efficient Robotic Manipulation", "sourceVersion": "2610.02120v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.191572+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.02120v1" }, "originalAnalysis": { "id": "arxiv-2610.02120", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458216+00:00", "sourceUrl": "https://arxiv.org/html/2610.02120v1", "sectionsRead": [ "3.1–3.5", "4.1–4.4", "Tables 1–4", "5 Conclusion" ], "methodsZh": "冻结视觉网络从RGB-D提取物体中心及交互点，正运动学提供机器人节点，组成统一三维骨架。两个Transformer专家以流匹配联合学习动作和未来骨架，后者仅作训练监督。推理省去未来分支，MAC从多条采样轨迹中选距离其他轨迹最近的一条，避免直接平均不兼容动作。", "experimentsZh": "在含10,030变体、七类扰动的完整LIBERO-Plus测试零样本成功率。实机ARX R5配外置及腕部RealSense，相同五任务每方法各20次，对比Fast-WAM、π0.5、Cosmos-Policy；另做骨架、监督和推理消融。", "resultsZh": "作者报告57.1M参数模型仿真85.9%，实机五任务均值89%；删去未来骨架监督降至80.1%，去掉MAC为84.2%。特权仿真状态参照为87.7%，不能混作纯视觉结果。", "limitationsZh": "布局扰动仅66.6%，低于π0.5的84.1%；真机样本有限，未证明复杂空间组合泛化。骨架依赖冻结感知和稳定节点对应，参数量优势不直接等于完整系统延迟优势。", "reproductionZh": "论文给出RTX4090训练60K步、有效批48，预测32步动作及八个未来骨架；推理十步积分、执行16步再观测，MAC取三候选前十步。完整复现仍需感知权重、标定与训练数据。", "experimentType": "both", "robots": [ "ARX R5" ], "evidenceNotes": [ { "section": "4.1; Table 1", "note": "LIBERO-Plus完整10,030变体；sim-state为特权输入。" }, { "section": "4.3; Table 2; Figure 4", "note": "ARX R5真机五任务各20次，均值89%。" }, { "section": "4.4; Tables 3–4", "note": "未来监督和MAC贡献分别通过80.1%、84.2%消融检验。" }, { "section": "PDF p6, Figure 4; Table 1", "note": "已渲染核对ARX R5标识及观测输入/特权状态两行，未把相机作机器人型号。" } ], "corrections": [ "experimentType从unknown修正为both；robots补充ARX R5；原摘要解读漏记真机实验。" ], "sourceVersion": "2610.02120v1", "originalSha256": "2518572d4baff871b3667fb5d22fd171f89ee6e87f192697a4e86f872adc4e49" }, "experimentNote": "在含10,030变体、七类扰动的完整LIBERO-Plus测试零样本成功率。实机ARX R5配外置及腕部RealSense，相同五任务每方法各20次，对比Fast-WAM、π0.5、Cosmos-Policy；另做骨架、监督和推理消融。", "contribution": "冻结视觉网络从RGB-D提取物体中心及交互点，正运动学提供机器人节点，组成统一三维骨架。两个Transformer专家以流匹配联合学习动作和未来骨架，后者仅作训练监督。推理省去未来分支，MAC从多条采样轨迹中选距离其他轨迹最近的一条，避免直接平均不兼容动作。", "whyUseful": "论文给出RTX4090训练60K步、有效批48，预测32步动作及八个未来骨架；推理十步积分、执行16步再观测，MAC取三候选前十步。完整复现仍需感知权重、标定与训练数据。", "analysisVerifiedAt": "2026-10-04T13:45:34.458216+00:00" }, { "id": "arxiv-2610.02161", "title": "DuoMind: Enabling Distributed Multi-Robot Coordination with Semantic Communication", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.02161v1", "titleZh": "DuoMind：通过语义通信实现分布式多机器人协调", "abstractZh": "每台机器人用Qwen3-VL-4B调度，用微调π0.5执行；上层交换意图、子目标、任务信念及可选不确定性。语言指令被限制在预设可接受集合内，并有停滞后强制推进机制，因此并非完全开放式自主协商。\n全部是仿真：ManiSkill3上RoboPoly七任务采用两台Panda；RoboTwin八任务把Aloha-AgileX双臂拆成独立代理。每任务50条双机示范拆成100条单机轨迹，每方法每任务400次评估。", "summary": "每台机器人用Qwen3-VL-4B调度，用微调π0.5执行；上层交换意图、子目标、任务信念及可选不确定性。语言指令被限制在预设可接受集合内，并有停滞后强制推进机制，因此并非完全开放式自主协商。", "experimentType": "sim", "robots": [ "Franka Emika Panda（仿真）", "Aloha-AgileX（仿真）" ], "tags": [ "多机器人协作", "语义通信", "视觉语言动作模型", "分层控制" ], "limitations": "只验证双代理及模拟桌面场景；共享全局相机，不能等同完全局部视觉。上层调用延迟、网络丢包及更大团队未得到实机验证；低成功任务说明规划改善仍受动作模型限制。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.02161v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.02161v1", "note": "3.2; 4; 4.1：平台、模拟环境、示范数量、400回合及指令限制。" }, { "url": "https://arxiv.org/html/2610.02161v1", "note": "Tables 1–2; 4.3–4.4：具体成功率及通信/动作模型消融。" }, { "url": "https://arxiv.org/pdf/2610.02161v1", "note": "3.2; 4; 4.1：平台、模拟环境、示范数量、400回合及指令限制。" }, { "url": "https://arxiv.org/pdf/2610.02161v1", "note": "Tables 1–2; 4.3–4.4：具体成功率及通信/动作模型消融。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "多机器人协作 / 语义通信", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "多机器人" ], "robotFilters": [ "Franka Panda", "Aloha" ], "original": { "id": "arxiv-2610.02161", "originalSourceUrl": "https://arxiv.org/pdf/2610.02161v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.02161v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.02161v1", "sha256": "058e0c6846bc17821a0fb12190f7f61ec3a5f35a0a290d3c12cf546f7aa4aee2", "bytes": 10966437, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:32:39.555501+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.02161v1", "license": "http://creativecommons.org/licenses/by-nc-sa/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.02161v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.02161v1", "pages": 19, "sourceIdentityStatus": "title_match", "sourceTitle": "DuoMind: Enabling Distributed Multi-Robot Coordination with Semantic Communication", "sourceVersion": "2610.02161v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.186553+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.02161v1" }, "originalAnalysis": { "id": "arxiv-2610.02161", "sourceUrl": "https://arxiv.org/pdf/2610.02161v1", "sectionsRead": [ "2.1–2.3", "3.1–3.3", "4.1–4.4", "Tables 1–2" ], "methodsZh": "每台机器人用Qwen3-VL-4B调度，用微调π0.5执行；上层交换意图、子目标、任务信念及可选不确定性。语言指令被限制在预设可接受集合内，并有停滞后强制推进机制，因此并非完全开放式自主协商。", "experimentsZh": "全部是仿真：ManiSkill3上RoboPoly七任务采用两台Panda；RoboTwin八任务把Aloha-AgileX双臂拆成独立代理。每任务50条双机示范拆成100条单机轨迹，每方法每任务400次评估。", "resultsZh": "RoboPoly的Cook Pot由π0.5单独控制的1.00%升至39.25%，Hang Bag从52.25%至78%；但Prepare Snack仅18%。RoboTwin多数任务改善，Put Object Cabinet反而由46.50%降到43.25%。去通信或换弱动作模型均降低总体效果。", "limitationsZh": "只验证双代理及模拟桌面场景；共享全局相机，不能等同完全局部视觉。上层调用延迟、网络丢包及更大团队未得到实机验证；低成功任务说明规划改善仍受动作模型限制。", "reproductionZh": "应保存上层提示词、可接受指令集、推进规则与LoRA配置，复用相同随机初态比较无通信、无调度两类消融，并单列RoboTwin训练高层指令、测试低层指令的分布差异。", "experimentType": "sim", "robots": [ "Franka Emika Panda（仿真）", "Aloha-AgileX（仿真）" ], "corrections": [ "实验类型由unknown修正为sim；论文没有真实多机器人部署结果。" ], "evidenceNotes": [ { "note": "平台、模拟环境、示范数量、400回合及指令限制。", "section": "3.2; 4; 4.1" }, { "note": "具体成功率及通信/动作模型消融。", "section": "Tables 1–2; 4.3–4.4" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2610.02161v1", "originalSha256": "058e0c6846bc17821a0fb12190f7f61ec3a5f35a0a290d3c12cf546f7aa4aee2" }, "experimentNote": "全部是仿真：ManiSkill3上RoboPoly七任务采用两台Panda；RoboTwin八任务把Aloha-AgileX双臂拆成独立代理。每任务50条双机示范拆成100条单机轨迹，每方法每任务400次评估。", "contribution": "每台机器人用Qwen3-VL-4B调度，用微调π0.5执行；上层交换意图、子目标、任务信念及可选不确定性。语言指令被限制在预设可接受集合内，并有停滞后强制推进机制，因此并非完全开放式自主协商。", "whyUseful": "应保存上层提示词、可接受指令集、推进规则与LoRA配置，复用相同随机初态比较无通信、无调度两类消融，并单列RoboTwin训练高层指令、测试低层指令的分布差异。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2610.02170", "title": "Watch, Infer, Coordinate: Inferring Robot Partner Constraints for Zero-Shot Coordination", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.02170v1", "titleZh": "观察、推断、协调：推断机器人伙伴的能力约束以实现零样本协作", "abstractZh": "把伙伴能力表示为有限候选约束，利用双方联合动作在有界理性行为模型下的似然形成后验；复杂场景以离线学习评分近似计算。部署时将推断约束加入CEM模型预测控制，预测受限伙伴对辅助者动作的响应。\n三类实验均为仿真：二维抬杆、固定基座双UR5及移动双UR5。以1、2、4、8段无障碍演示推断约束，再用8段演示在新障碍任务中评测；同一规划器对比未知能力、真值能力、CE-CM-Div及行为克隆，报告三随机种子的均值与标准误。", "summary": "把伙伴能力表示为有限候选约束，利用双方联合动作在有界理性行为模型下的似然形成后验；复杂场景以离线学习评分近似计算。部署时将推断约束加入CEM模型预测控制，预测受限伙伴对辅助者动作的响应。", "experimentType": "sim", "robots": [ "UR5（仿真）" ], "tags": [ "多机器人协作", "能力推断", "零样本泛化", "操作基准" ], "limitations": "假定已知目标、动力学及伙伴行为模型，状态动作完全可观测，真实约束属于有限候选集；未验证实机、未知伙伴策略或约束随时间变化。行为克隆失败不代表所有学习型协作方法均无效。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.02170v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.02170v1", "note": "§3.3, §5.1：三个仿真设置及演示预算、三随机种子协议。" }, { "url": "https://arxiv.org/html/2610.02170v1", "note": "Tables 2–3; §6：约束误差、成功率和已知模型假设。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "多机器人协作 / 能力推断", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "操作与抓取", "数据集与基准", "多机器人" ], "robotFilters": [ "Universal Robots UR5" ], "original": { "id": "arxiv-2610.02170", "originalSourceUrl": "https://arxiv.org/pdf/2610.02170v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.02170v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.02170v1", "sha256": "79d41ca531a2b8df3f7f935ba455bc2445ff52ba493dcce5fd8dbe6d61eec045", "bytes": 5929632, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:32:28.852783+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.02170v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.02170v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.02170v1", "pages": 17, "sourceIdentityStatus": "title_match", "sourceTitle": "Watch, Infer, Coordinate: Inferring Robot Partner Constraints for Zero-Shot Coordination", "sourceVersion": "2610.02170v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.173482+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.02170v1" }, "originalAnalysis": { "id": "arxiv-2610.02170", "sourceUrl": "https://arxiv.org/html/2610.02170v1", "sectionsRead": [ "§3–4", "§5 Tables 2–3", "§6" ], "methodsZh": "把伙伴能力表示为有限候选约束，利用双方联合动作在有界理性行为模型下的似然形成后验；复杂场景以离线学习评分近似计算。部署时将推断约束加入CEM模型预测控制，预测受限伙伴对辅助者动作的响应。", "experimentsZh": "三类实验均为仿真：二维抬杆、固定基座双UR5及移动双UR5。以1、2、4、8段无障碍演示推断约束，再用8段演示在新障碍任务中评测；同一规划器对比未知能力、真值能力、CE-CM-Div及行为克隆，报告三随机种子的均值与标准误。", "resultsZh": "作者报告三场景新任务成功率92.6%、85.2%、63.0%，真值参照为96.3%、88.9%、70.4%。移动双臂约束汉明误差随演示从1段增至8段由0.190降至0.009；联合观察通常优于只看受限机器人。", "limitationsZh": "假定已知目标、动力学及伙伴行为模型，状态动作完全可观测，真实约束属于有限候选集；未验证实机、未知伙伴策略或约束随时间变化。行为克隆失败不代表所有学习型协作方法均无效。", "reproductionZh": "复现需同时固定演示任务、候选能力集、行为温度、MPC预算和随机种子；先比较约束误差再比较成功率及SNA。文中承诺发布代码，本分析未运行实现。", "experimentType": "sim", "robots": [ "UR5（仿真）" ], "evidenceNotes": [ { "section/page": "§3.3, §5.1", "note": "三个仿真设置及演示预算、三随机种子协议。" }, { "section/page": "Tables 2–3; §6", "note": "约束误差、成功率和已知模型假设。" } ], "corrections": [ { "field": "experimentType", "value": "sim", "reason": "正文明确仅仿真；UR5不是实机证据。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04" }, "experimentNote": "三类实验均为仿真：二维抬杆、固定基座双UR5及移动双UR5。以1、2、4、8段无障碍演示推断约束，再用8段演示在新障碍任务中评测；同一规划器对比未知能力、真值能力、CE-CM-Div及行为克隆，报告三随机种子的均值与标准误。", "contribution": "把伙伴能力表示为有限候选约束，利用双方联合动作在有界理性行为模型下的似然形成后验；复杂场景以离线学习评分近似计算。部署时将推断约束加入CEM模型预测控制，预测受限伙伴对辅助者动作的响应。", "whyUseful": "复现需同时固定演示任务、候选能力集、行为温度、MPC预算和随机种子；先比较约束误差再比较成功率及SNA。文中承诺发布代码，本分析未运行实现。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2610.02196", "title": "InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.02196v1", "titleZh": "InterEvolve：面向人形机器人移动操作的测试时奖励程序演化", "abstractZh": "以分阶段奖励、结束条件和可调常数组成程序；语言模型修改结构，CMA-ES校准参数，固定验证器比较并行仿真结果。物体残差加在冻结的身体前向—后向模型上，使奖励映射成潜变量，执行期间不再训练控制器。\n在Isaac Lab中使用重定向到G1的OMOMO/GRAB交互数据；四类箱状物各留出50片段。比较参考跟踪、八类目标任务和技能库组合，指标取三次独立评估；真机G1通过胸前RGB-D和LiDAR估计物体。", "summary": "以分阶段奖励、结束条件和可调常数组成程序；语言模型修改结构，CMA-ES校准参数，固定验证器比较并行仿真结果。物体残差加在冻结的身体前向—后向模型上，使奖励映射成潜变量，执行期间不再训练控制器。", "experimentType": "both", "robots": [ "Unitree G1" ], "tags": [ "人形机器人", "移动操作", "奖励设计", "测试时适应" ], "limitations": "能力受已有运动库与状态采样库限制；搜索耗时不支持实机实时重规划。机载传感器不等于机载计算：感知和策略在机外GPU执行，真机回放仿真潜变量，并保留状态反馈闭环。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.02196v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.02196v1", "note": "Table 2：八任务族平均SR 86.5%；非实机成功率。" }, { "url": "https://arxiv.org/html/2610.02196v1", "note": "Appendix D.8：机外GPU工作站；回放仿真潜变量，控制器以本体和物体估计闭环。" }, { "url": "https://arxiv.org/html/2610.02196v1", "note": "Appendix E：搜索需要GPU小时，不能实时物理重规划。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "人形机器人 / 移动操作", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "操作与抓取", "人形机器人" ], "robotFilters": [ "Unitree G1" ], "original": { "id": "arxiv-2610.02196", "originalSourceUrl": "https://arxiv.org/pdf/2610.02196v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.02196v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.02196v1", "sha256": "37e61f9f5a3b7daa9d09ced06b8e72537b4815afe1cc4ee1f1a10c0bf897542b", "bytes": 11868445, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:32:29.060061+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.02196v1", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.02196v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.02196v1", "pages": 31, "sourceIdentityStatus": "title_match", "sourceTitle": "InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation", "sourceVersion": "2610.02196v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.164776+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.02196v1" }, "originalAnalysis": { "id": "arxiv-2610.02196", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090087+00:00", "sourceUrl": "https://arxiv.org/html/2610.02196v1", "sectionsRead": [ "§3", "§4.1–4.4", "Tables 1–2,16", "Appendix D.6–D.8", "Appendix E" ], "methodsZh": "以分阶段奖励、结束条件和可调常数组成程序；语言模型修改结构，CMA-ES校准参数，固定验证器比较并行仿真结果。物体残差加在冻结的身体前向—后向模型上，使奖励映射成潜变量，执行期间不再训练控制器。", "experimentsZh": "在Isaac Lab中使用重定向到G1的OMOMO/GRAB交互数据；四类箱状物各留出50片段。比较参考跟踪、八类目标任务和技能库组合，指标取三次独立评估；真机G1通过胸前RGB-D和LiDAR估计物体。", "resultsZh": "作者表2报告目标任务成功率86.5%，校准后的初始语言模型程序为34.6%；每任务族约2.1 GPU小时、23万token。跟踪成功率72%，但专门跟踪器的轨迹误差更低，不能说全部指标领先。", "limitationsZh": "能力受已有运动库与状态采样库限制；搜索耗时不支持实机实时重规划。机载传感器不等于机载计算：感知和策略在机外GPU执行，真机回放仿真潜变量，并保留状态反馈闭环。", "reproductionZh": "复现需固定训练/留出划分、奖励特征、验证器和搜索预算，区分单次回合成功与三次尝试中成功的场景口径；论文列明流程，但本轮未运行代码或重现实验。", "experimentType": "both", "robots": [ "Unitree G1" ], "evidenceNotes": [ { "section/page": "Table 2", "note": "八任务族平均SR 86.5%；非实机成功率。" }, { "section/page": "Appendix D.8", "note": "机外GPU工作站；回放仿真潜变量，控制器以本体和物体估计闭环。" }, { "section/page": "Appendix E", "note": "搜索需要GPU小时，不能实时物理重规划。" } ], "corrections": { "deploymentNote": "机载感知输入、机外GPU计算；真实执行不是实机在线奖励演化。" } }, "experimentNote": "在Isaac Lab中使用重定向到G1的OMOMO/GRAB交互数据；四类箱状物各留出50片段。比较参考跟踪、八类目标任务和技能库组合，指标取三次独立评估；真机G1通过胸前RGB-D和LiDAR估计物体。", "contribution": "以分阶段奖励、结束条件和可调常数组成程序；语言模型修改结构，CMA-ES校准参数，固定验证器比较并行仿真结果。物体残差加在冻结的身体前向—后向模型上，使奖励映射成潜变量，执行期间不再训练控制器。", "whyUseful": "复现需固定训练/留出划分、奖励特征、验证器和搜索预算，区分单次回合成功与三次尝试中成功的场景口径；论文列明流程，但本轮未运行代码或重现实验。", "analysisVerifiedAt": "2026-10-04T13:50:35.090087+00:00" }, { "id": "arxiv-2610.02204", "title": "Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents", "date": "2026-10-01", "url": "https://arxiv.org/abs/2610.02204v1", "titleZh": "重建、练习、走向真实：面向具身智能体的引导式自我改进", "abstractZh": "RPG不更新模型权重，而用离线视频构建MuJoCo练习任务。执行智能体与可见特权状态的智能体并行尝试，视频分析定位失败，再修改提示词和共享技能。候选及合并版本须提高跨任务均值，且单任务下降不超过五次中的一次，才可保留。\n22项仿真任务以每项五个固定开发种子迭代15轮，冻结后在每项十个留出种子评价。实机YAM双臂做收球关抽屉、折毛巾、转交碗，各十次；共同标定后不按任务调参，每次最多30次模型调用或20分钟，允许自主重试。", "summary": "RPG不更新模型权重，而用离线视频构建MuJoCo练习任务。执行智能体与可见特权状态的智能体并行尝试，视频分析定位失败，再修改提示词和共享技能。候选及合并版本须提高跨任务均值，且单任务下降不超过五次中的一次，才…", "experimentType": "both", "robots": [ "YAM" ], "tags": [ "具身智能体", "自主技能学习", "仿真到现实", "大语言模型" ], "limitations": "改进曲线只有一次开发运行，轮次不等算力预算。仿真11次失败中七次来自折毛巾；实机折叠判据较宽且独立定义。额外零样本任务仅定性展示，小规模30次试验不能证明通用可靠性。", "translationType": "中文原文分析（非逐字全文翻译）", "tier": "recent", "status": "训练代码待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2610.02204v1", "note": "依据原文摘要；实验范围与型号待全文复核" }, { "url": "https://arxiv.org/html/2610.02204v1", "note": "III-D; IV-A：每候选110次开发评测；最终每系统220次留出评测。" }, { "url": "https://arxiv.org/html/2610.02204v1", "note": "IV-E; Table V：物理YAM三任务各十次，允许试次内自主恢复。" }, { "url": "https://arxiv.org/html/2610.02204v1", "note": "IV-B; Figure 4：单次改进运行，计算预算不匹配；实机和仿真折叠判据不同。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "具身智能体 / 自主技能学习", "year": 2026, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "仅核对摘要；官方实现、许可证与训练入口尚待核实。", "directions": [ "其他机器人研究" ], "robotFilters": [ "YAM" ], "original": { "id": "arxiv-2610.02204", "originalSourceUrl": "https://arxiv.org/pdf/2610.02204v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.02204v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.02204v1", "sha256": "93feecda4aa9f514398bf49f8e6494173cbfee6a6f6ca5e34520b1dd46efbffa", "bytes": 8626170, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:32:28.852010+00:00", "publicationDate": "2026/10/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.02204v1", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.02204v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.02204v1", "pages": 17, "sourceIdentityStatus": "title_match", "sourceTitle": "Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents", "sourceVersion": "2610.02204v1", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.151338+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.02204v1" }, "originalAnalysis": { "id": "arxiv-2610.02204", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458204+00:00", "sourceUrl": "https://arxiv.org/html/2610.02204v1", "sectionsRead": [ "III-A–E", "IV-A–E", "Tables II–V", "Figures 4–6" ], "methodsZh": "RPG不更新模型权重，而用离线视频构建MuJoCo练习任务。执行智能体与可见特权状态的智能体并行尝试，视频分析定位失败，再修改提示词和共享技能。候选及合并版本须提高跨任务均值，且单任务下降不超过五次中的一次，才可保留。", "experimentsZh": "22项仿真任务以每项五个固定开发种子迭代15轮，冻结后在每项十个留出种子评价。实机YAM双臂做收球关抽屉、折毛巾、转交碗，各十次；共同标定后不按任务调参，每次最多30次模型调用或20分钟，允许自主重试。", "resultsZh": "作者报告仿真209/220成功，实机30/30；五轮消融中完整系统78.2%，去掉视频分析或特权智能体约51%。固定运行时、只改技能库的四任务对照由26/40升至37/40。", "limitationsZh": "改进曲线只有一次开发运行，轮次不等算力预算。仿真11次失败中七次来自折毛巾；实机折叠判据较宽且独立定义。额外零样本任务仅定性展示，小规模30次试验不能证明通用可靠性。", "reproductionZh": "复现需保存任务判据、开发/测试种子、完整技能快照、模型版本及调用预算，并实施YAM坐标与夹爪统一标定；论文分析不等于已运行代码或验证硬件安全。", "experimentType": "both", "robots": [ "YAM" ], "evidenceNotes": [ { "section": "III-D; IV-A", "note": "每候选110次开发评测；最终每系统220次留出评测。" }, { "section": "IV-E; Table V", "note": "物理YAM三任务各十次，允许试次内自主恢复。" }, { "section": "IV-B; Figure 4", "note": "单次改进运行，计算预算不匹配；实机和仿真折叠判据不同。" } ], "corrections": [ "robots由未核实补充为YAM；30/30明确为含自主重试的完整试次结果。" ], "sourceVersion": "2610.02204v1", "originalSha256": "93feecda4aa9f514398bf49f8e6494173cbfee6a6f6ca5e34520b1dd46efbffa" }, "experimentNote": "22项仿真任务以每项五个固定开发种子迭代15轮，冻结后在每项十个留出种子评价。实机YAM双臂做收球关抽屉、折毛巾、转交碗，各十次；共同标定后不按任务调参，每次最多30次模型调用或20分钟，允许自主重试。", "contribution": "RPG不更新模型权重，而用离线视频构建MuJoCo练习任务。执行智能体与可见特权状态的智能体并行尝试，视频分析定位失败，再修改提示词和共享技能。候选及合并版本须提高跨任务均值，且单任务下降不超过五次中的一次，才可保留。", "whyUseful": "复现需保存任务判据、开发/测试种子、完整技能快照、模型版本及调用预算，并实施YAM坐标与夹爪统一标定；论文分析不等于已运行代码或验证硬件安全。", "analysisVerifiedAt": "2026-10-04T13:45:34.458204+00:00" }, { "id": "arxiv-2609.40208", "title": "Centralized Multi-UAV Exploration and 3D Reconstruction Using Single-UAV Planners", "titleZh": "利用单无人机规划器实现集中式多无人机探索与三维重建", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2609.40208", "paperUrl": "https://arxiv.org/abs/2609.40208", "projectUrl": null, "codeUrl": null, "category": "多机器人协同与探索", "tags": [ "多无人机", "TSDF融合", "集中规划", "仿真" ], "directions": [ "导航与建图" ], "robotFilters": [ "DJI F450" ], "tier": "recent", "summary": "通过共享地图、路径避碰和机体过滤复用四种单机探索规划器；三机仿真中分散起点通常更快，无实体无人机实验。", "abstractZh": "该工作主要贡献系统集成：每架无人机局部融合深度点云，中央服务器合并增量TSDF并为各机异步规划，结合共享路径和其他机体位置避免碰撞及把同伴重建为静态障碍。实验比较一起出发与空间分散出发，显示初始部署影响覆盖效率；没有新增显式任务分配算法或真机验证。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "MRS仿真三架DJI F450配D435i，在迷宫、警局和学校环境比较。迷宫只测分散起点，另两环境比较联合/分散起点；每条件五次。模拟RTK-GPS提供一致定位，工作站同时承担物理仿真、多机局部融合及中央处理。", "robots": [ "三架DJI F450（MRS仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现要固定地图分辨率、最小避障距离、起始区域及同步方式，并分清迷宫秒与其他场景分钟单位。路径长度是三机总和而速度为各机均值；应记录未达95%的删失回合，而非当成零时长。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "保留RH-NBVP、KRH-NBVP、AEP、KAEP核心采样逻辑，在机端维护局部Voxblox地图并传增量，中央按权重融合TSDF。各规划器访问全局图及活动路径缓存，剔除靠近障碍/同伴路线的候选；点云滤掉其他机体周围球形区域。", "whyUseful": "复现要固定地图分辨率、最小避障距离、起始区域及同步方式，并分清迷宫秒与其他场景分钟单位。路径长度是三机总和而速度为各机均值；应记录未达95%的删失回合，而非当成零时长。", "limitations": "只验证三机与可靠定位假设，尚未证明实际通信丢包、时延、漂移下安全性或大队伍扩展。分散部署不一定更省总航程；迷宫缺联合起点对照，不能声称三个场景都直接验证了该对比。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2609.40208v1", "note": "PDF第4–5页§IV：明确simulation，三架DJI F450及模拟RTK-GPS；未报告真机。" }, { "url": "https://arxiv.org/pdf/2609.40208v1", "note": "PDF第7页表III（已渲染核看）：迷宫单位秒，其余分钟；横线表示时间限制内没达到95%。" }, { "url": "https://arxiv.org/pdf/2609.40208v1", "note": "PDF第5页部署说明：迷宫仅SS，警局和学校同时有JS/SS；不存在迷宫JS直接比较。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T14:02:45.219445+00:00", "original": { "id": "arxiv-2609.40208", "originalSourceUrl": "https://arxiv.org/pdf/2609.40208v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.40208v1", "finalDownloadUrl": "https://arxiv.org/pdf/2609.40208", "sha256": "34d760ade1448af94a2e1f5dee7a285f94ca458fc30a38a9f8fa314874733f4d", "bytes": 4137949, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:45:00.353255+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.40208", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.40208", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.40208", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "Centralized Multi-UAV Exploration and 3D Reconstruction Using Single-UAV Planners", "sourceVersion": "2609.40208v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.830323+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.40208" }, "originalAnalysis": { "id": "arxiv-2609.40208", "sectionsRead": [ "第III节系统架构与算法，PDF第3–4页", "第IV节实验与表I–III，PDF第4–7页", "第V节结论，PDF第6–7页（选段）" ], "analyzedAt": "2026-10-04T14:02:45.219445+00:00", "analysisStatus": "full_text_sections", "methodsZh": "保留RH-NBVP、KRH-NBVP、AEP、KAEP核心采样逻辑，在机端维护局部Voxblox地图并传增量，中央按权重融合TSDF。各规划器访问全局图及活动路径缓存，剔除靠近障碍/同伴路线的候选；点云滤掉其他机体周围球形区域。", "experimentsZh": "MRS仿真三架DJI F450配D435i，在迷宫、警局和学校环境比较。迷宫只测分散起点，另两环境比较联合/分散起点；每条件五次。模拟RTK-GPS提供一致定位，工作站同时承担物理仿真、多机局部融合及中央处理。", "resultsZh": "KAEP在学校95%覆盖由联合起点5.86分钟变为分散5.35分钟，警局由2.96降至2.60分钟；迷宫分散起点59.04秒达95%。学校中央TSDF合并8.6毫秒、地图RMSE0.19米。RH-NBVP若干配置仍未在限时内达到95%。", "limitationsZh": "只验证三机与可靠定位假设，尚未证明实际通信丢包、时延、漂移下安全性或大队伍扩展。分散部署不一定更省总航程；迷宫缺联合起点对照，不能声称三个场景都直接验证了该对比。", "reproductionZh": "复现要固定地图分辨率、最小避障距离、起始区域及同步方式，并分清迷宫秒与其他场景分钟单位。路径长度是三机总和而速度为各机均值；应记录未达95%的删失回合，而非当成零时长。", "sourceUrl": "https://arxiv.org/pdf/2609.40208v1", "sourceVersion": "2609.40208v1", "originalSha256": "34d760ade1448af94a2e1f5dee7a285f94ca458fc30a38a9f8fa314874733f4d", "experimentType": "sim", "robots": [ "三架DJI F450（MRS仿真）" ], "evidenceNotes": [ { "section": "PDF第4–5页§IV", "note": "明确simulation，三架DJI F450及模拟RTK-GPS；未报告真机。" }, { "section": "PDF第7页表III（已渲染核看）", "note": "迷宫单位秒，其余分钟；横线表示时间限制内没达到95%。" }, { "section": "PDF第5页部署说明", "note": "迷宫仅SS，警局和学校同时有JS/SS；不存在迷宫JS直接比较。" } ], "verification": "verified", "directionsOriginal": [ "协同探索", "三维重建" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Extending single Unmanned Aerial Vehicles (UAVs) exploration methods to multi-UAV teams can improve coverage speed and robustness, but introduces challenges such as consistent mapping, safe navigation, and deployment strategy. In this work, we present a centralized multi-UAV exploration framework that enables the use of existing single-UAV sampling-based planners in a multi-UAV setting. The proposed architecture allows multiple UAVs to collaboratively explore unknown environments using a shared global Truncated Signed Distance Field (TSDF) map and centralized planning. Building on the voxblox library, we adapt its mapping pipeline to support real-time fusion of depth measurements from multiple UAVs into a common TSDF representation. In addition, inter-UAV collision avoidance and robot self-filtering mechanisms are integrated into the system to ensure safe navigation and prevent reconstruction of other UAVs as static obstacles. The framework is evaluated in simulation using four sampling-based exploration planners - RH-NBVP, KRH-NBVP, AEP, and KAEP - whose core sampling logic is preserved, with only system-level adaptations for multi-UAV operation. Experiments are conducted across multiple environments and under two deployment configurations: Joint Start (JS), where UAVs are initialized in close proximity, and Separated Start (SS), where UAVs are initialized in distinct locations. Results show that SS deployments consistently achieve faster exploration and improved coverage across all planners, highlighting the importance of the deployment strategy in multi-UAV exploration performance." }, { "id": "arxiv-2609.40244", "title": "StreamRig: Exploiting Intra-Rig Geometry for Streaming Multi-Camera Odometry", "titleZh": "StreamRig：利用相机组内几何实现流式多相机里程计", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2609.40244", "paperUrl": "https://arxiv.org/abs/2609.40244", "projectUrl": null, "codeUrl": null, "category": "定位与三维感知", "tags": [ "多相机", "流式里程计", "冻结基础模型", "仿真到真实" ], "directions": [ "导航与建图" ], "robotFilters": [ "Segway", "KITTI-360采集车辆" ], "tier": "recent", "summary": "冻结多视图几何模型并训练紧凑时序模块，四数据集相对漂移改善；真实机器人序列评测不等于闭环导航执行。", "abstractZh": "StreamRig把已标定多相机图像作为一个几何观测，用少量潜词元连接冻结的多视图基础模型与因果时序估计器。分组重定位预训练提供几何对齐初始化，再学习带锚点快照的流式位姿回归，定期换锚限制状态规模。实验包含真实机器人采集数据，但任务是里程计估计与轨迹评分。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "data", "experimentNote": "NCLT、TartanGround、KITTI-360及ZJH四组数据评测完整因果轨迹；ZJH用131条仿真人形步行轨迹训练，在三条真实序列零样本测试。另有相机数量、训练窗口、锚点、归一化及前端替换消融。", "robots": [ "NCLT Segway（数据采集）", "TartanGround地面机器人（仿真数据）", "KITTI-360车辆（数据采集）", "ZJH人形机器人（真实序列，型号未列）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现需分别固定采样步长、SE(3)/Sim(3)对齐及段长，排除获得真实相对位姿的oracle。准确率与延迟采用不同换锚配置，不能合并成单一速度精度工作点；真实序列仅离线位姿评测，无导航控制实验。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "冻结MapAnything同时编码已知内外参和同步多视图，Rig-Resampler每相机压成16词元。八层CausalBridge缓存历史键值，各查询拥有独立锚点副本；只训练74.6M后端参数，以相对位姿监督，先分组重定位再因果序列适配。", "whyUseful": "复现需分别固定采样步长、SE(3)/Sim(3)对齐及段长，排除获得真实相对位姿的oracle。准确率与延迟采用不同换锚配置，不能合并成单一速度精度工作点；真实序列仅离线位姿评测，无导航控制实验。", "limitations": "ZJH所有方法都用整段真值估计一个尺度，不代表直接恢复准确公制尺度。零样本仿真到真实只针对ZJH，其他数据集使用目标训练集。长锚间隔超出训练位移范围时误差上升，闭环检测和地图约束尚未定量验证。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2609.40244v1", "note": "PDF第5页表I（已渲染核看）：ZJH统一up-to-scale；KITTI-360的STream3R ATE53.6低于本法63.7。" }, { "url": "https://arxiv.org/pdf/2609.40244v1", "note": "PDF第4页§IV-A：只有ZJH明确纯仿真训练、三真实序列零样本；NCLT/KITTI均用自身训练划分。" }, { "url": "https://arxiv.org/pdf/2609.40244v1", "note": "PDF第7页§IV-E：图1精度N=24、计算开销N=2分开测量；bounded memory来自重新锚定。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T14:02:45.217740+00:00", "original": { "id": "arxiv-2609.40244", "originalSourceUrl": "https://arxiv.org/pdf/2609.40244v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.40244v1", "finalDownloadUrl": "https://arxiv.org/pdf/2609.40244", "sha256": "4517b0d05796f974a8056133a25059a891b65f1d9bf244932c45195a37175358", "bytes": 3233792, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:44:50.641305+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.40244", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.40244", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.40244", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "StreamRig: Exploiting Intra-Rig Geometry for Streaming Multi-Camera Odometry", "sourceVersion": "2609.40244v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.778011+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.40244" }, "originalAnalysis": { "id": "arxiv-2609.40244", "sectionsRead": [ "第III节方法，PDF第2–4页", "第IV节实验及表I–V，PDF第4–8页", "第V节结论，PDF第8页" ], "analyzedAt": "2026-10-04T14:02:45.217740+00:00", "analysisStatus": "full_text_sections", "methodsZh": "冻结MapAnything同时编码已知内外参和同步多视图，Rig-Resampler每相机压成16词元。八层CausalBridge缓存历史键值，各查询拥有独立锚点副本；只训练74.6M后端参数，以相对位姿监督，先分组重定位再因果序列适配。", "experimentsZh": "NCLT、TartanGround、KITTI-360及ZJH四组数据评测完整因果轨迹；ZJH用131条仿真人形步行轨迹训练，在三条真实序列零样本测试。另有相机数量、训练窗口、锚点、归一化及前端替换消融。", "resultsZh": "四组非oracle方法中平移/旋转相对漂移最低；NCLT为2.77%和1.39°/100米，ZJH为3.48%和14.88°/100米。五相机输入推理26.2毫秒、2.6 GiB；KITTI-360的ATE不是所有方法最优，指标不能混同。", "limitationsZh": "ZJH所有方法都用整段真值估计一个尺度，不代表直接恢复准确公制尺度。零样本仿真到真实只针对ZJH，其他数据集使用目标训练集。长锚间隔超出训练位移范围时误差上升，闭环检测和地图约束尚未定量验证。", "reproductionZh": "复现需分别固定采样步长、SE(3)/Sim(3)对齐及段长，排除获得真实相对位姿的oracle。准确率与延迟采用不同换锚配置，不能合并成单一速度精度工作点；真实序列仅离线位姿评测，无导航控制实验。", "sourceUrl": "https://arxiv.org/pdf/2609.40244v1", "sourceVersion": "2609.40244v1", "originalSha256": "4517b0d05796f974a8056133a25059a891b65f1d9bf244932c45195a37175358", "experimentType": "data", "robots": [ "NCLT Segway（数据采集）", "TartanGround地面机器人（仿真数据）", "KITTI-360车辆（数据采集）", "ZJH人形机器人（真实序列，型号未列）" ], "evidenceNotes": [ { "section": "PDF第5页表I（已渲染核看）", "note": "ZJH统一up-to-scale；KITTI-360的STream3R ATE53.6低于本法63.7。" }, { "section": "PDF第4页§IV-A", "note": "只有ZJH明确纯仿真训练、三真实序列零样本；NCLT/KITTI均用自身训练划分。" }, { "section": "PDF第7页§IV-E", "note": "图1精度N=24、计算开销N=2分开测量；bounded memory来自重新锚定。" } ], "verification": "verified", "directionsOriginal": [ "视觉定位", "三维基础模型迁移" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Mobile robots and vehicles carry synchronized multi-camera rigs, yet many streaming 3D foundation models are designed for monocular input, leaving efficient use of rig geometry a challenge. We present StreamRig, a freeze-and-stream framework that builds causal streaming odometry for calibrated rigs on a frozen multi-view 3D foundation model. The frozen front-end jointly perceives the synchronized views using rig calibration. A Rig-Resampler compresses their features, a CausalBridge applies causal attention with a key-value cache, and a lightweight head regresses rig poses. A periodic re-anchoring protocol supports stable pose estimation over long sequences. Only these modules are trained, 74.6M parameters in total, with relative poses as the sole supervision. Our two-stage training strategy combines group relocalization pretraining with causal rig training to transfer the geometric priors of the frozen front-end and the alignment ability of the pretrained modules to streaming odometry. We evaluate on NCLT, TartanGround, KITTI-360, and our self-collected humanoid-robot dataset ZJH, where training uses only simulation and real-world evaluation is zero-shot. Across all four datasets, StreamRig achieves lower translation and rotation drift than the evaluated non-oracle monocular streaming and rig-aware offline models, while maintaining low inference cost. Ablations and controlled camera-count experiments identify the sources of these gains. We further examine how longer training windows affect inference over longer horizons. Code has been released at https://github.com/WeiYuFei0217/StreamRig." }, { "id": "arxiv-2609.40245", "title": "STARS: From Spatiotemporal Dynamics to Social Representations in Human-Robot Interaction", "titleZh": "STARS：从时空动力学到人机交互中的社会表征", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2609.40245", "paperUrl": "https://arxiv.org/abs/2609.40245", "projectUrl": null, "codeUrl": null, "category": "人机交互与社会导航", "tags": [ "图神经网络", "自监督表征", "社会导航", "离线评测" ], "directions": [], "robotFilters": [], "tier": "recent", "summary": "以时空图自编码学习社会交互表征，在离线行人动作分类上体现较强数据效率；原目录摘要与当前PDF内容错配。 来源警告：页面摘要与v2原文不一致，分析以PDF为准。", "abstractZh": "STARS把短时间窗内的人与机器人建成时空关系图，通过自监督图变分自编码器压缩成节点表示，再冻结表示训练线性探测器。其重点是从几何动态中学习可迁移社会信息，使用SEAN-T和SocialNav-SUB现有数据评估，而不是重新提出SocialNav-SUB视觉问答基准，也没有闭环机器人导航实验。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "data", "experimentNote": "联合使用SEAN-T的八秒VR导航观察与SNS行人交互数据，评测能力、意图、惊讶等主观感知以及行人关系动作分类。逐步使用1%至100%标签，比较MLP、自编码器、随机森林及同构/异构图，十个随机种子。", "robots": [ "SEAN-T虚拟导航代理", "SNS数据中的机器人与行人（离线轨迹）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现需固定图构建、有效节点掩码、标签比例、随机种子与冻结线性探测协议，避免把下游标签泄露到自监督阶段。本次以v2原文改正输入摘要错配，项目代码是否实际可下载未独立核验。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "用完整有向图表示同时观察到的代理，边包含相对SE(2)位姿及时间差分。异构版本区分人、机器人与关系类型；消息传递图VAE重建节点有效属性和时序边特征，以KL正则学习每节点潜变量，再冻结编码器作线性探测。", "whyUseful": "复现需固定图构建、有效节点掩码、标签比例、随机种子与冻结线性探测协议，避免把下游标签泄露到自监督阶段。本次以v2原文改正输入摘要错配，项目代码是否实际可下载未独立核验。", "limitations": "来源警告：arXiv落地页摘要与同标识v2 PDF不一致。落地页摘要描述SocialNav-SUB视觉问答基准，v2 PDF则是STARS时空图自监督表征研究；已确认PDF标题、作者、标识和版本匹配。本站以v2 PDF分析，未沿用误配摘要；上游不一致原因尚未确定。 评测依赖处理好的轨迹和固定窗口，未验证在线分段、感知噪声或闭环导航。图聚合与压缩会平滑精细几何，对群体检测等边级任务可能不利；潜空间可视化只是定性证据。", "caveats": "来源警告：arXiv落地页摘要与同标识v2 PDF不一致。落地页摘要描述SocialNav-SUB视觉问答基准，v2 PDF则是STARS时空图自监督表征研究；已确认PDF标题、作者、标识和版本匹配。本站以v2 PDF分析，未沿用误配摘要；上游不一致原因尚未确定。 性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2609.40245v2", "note": "来源警告：arXiv落地页摘要与同标识v2 PDF不一致。落地页摘要描述SocialNav-SUB视觉问答基准，v2 PDF则是STARS时空图自监督表征研究；已确认PDF标题、作者、标识和版本匹配。本站以v2 PDF分析，未沿用误配摘要；上游不一致原因尚未确定。" }, { "url": "https://arxiv.org/abs/2609.40245v2", "note": "已核实存在摘要不一致的arXiv版本页；只用于说明元数据冲突，不作为STARS方法摘要。" }, { "url": "https://arxiv.org/pdf/2609.40245v2", "note": "PDF第1页摘要：arXiv v2（2026-10-01）是STARS图表征方法；输入abstractOriginal为SocialNav-SUB基准摘要，二者错配。" }, { "url": "https://arxiv.org/pdf/2609.40245v2", "note": "PDF第8页表1（已渲染核看）：SNS宏F1=0.503；SEAN-T并非每个标注比例或指标都占优。" }, { "url": "https://arxiv.org/pdf/2609.40245v2", "note": "PDF第7页实验与第9页局限：基于既有交互数据做离线预测，没有执行机器人导航政策。" } ], "verification": "partial", "verificationNote": "来源警告：arXiv落地页摘要与同标识v2 PDF不一致。落地页摘要描述SocialNav-SUB视觉问答基准，v2 PDF则是STARS时空图自监督表征研究；已确认PDF标题、作者、标识和版本匹配。本站以v2 PDF分析，未沿用误配摘要；上游不一致原因尚未确定。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:53:31.184413+00:00", "original": { "id": "arxiv-2609.40245", "originalSourceUrl": "https://arxiv.org/pdf/2609.40245v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.40245v2", "finalDownloadUrl": "https://arxiv.org/pdf/2609.40245", "sha256": "9d87aa14858069890ac9bfcb61bfbe9884777d11f70eb7900fffe9e5adf650fc", "bytes": 3859822, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:44:50.774302+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.40245", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.40245", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked_with_abstract_conflict", "metadataSourceUrl": "https://arxiv.org/abs/2609.40245", "pages": 26, "sourceIdentityStatus": "title_and_arxiv_id_match_abstract_conflict", "sourceTitle": "STARS: From Spatiotemporal Dynamics to Social Representations in Human-Robot Interaction", "sourceVersion": "2609.40245v2", "sourceVersionDate": "2026/10/01", "archiveValidationStatus": "file_validated_metadata_conflict", "archiveValidatedAt": "2026-10-04T13:45:15.774332+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.40245", "metadataConflict": { "status": "confirmed_upstream_abstract_inconsistency", "affectedFields": [ "arXiv landing-page abstract", "arXiv citation_abstract", "collector abstractOriginal" ], "finding": "Current unversioned and versioned v2 arXiv landing metadata use the STARS title, but their abstract describes introducing the SocialNav-SUB VQA benchmark. The versioned v2 PDF instead presents STARS, a self-supervised spatiotemporal graph autoencoder with linear-probe evaluation. Its title, authors, arXiv ID, and version match the archive. This is an upstream metadata/PDF abstract inconsistency, not a mismatched local PDF. The cause of the upstream inconsistency is not established.", "currentLandingUrl": "https://arxiv.org/abs/2609.40245", "currentVersionedLandingUrl": "https://arxiv.org/abs/2609.40245v2", "versionedPdfUrl": "https://arxiv.org/pdf/2609.40245v2", "currentVersionedPdfSha256": "9d87aa14858069890ac9bfcb61bfbe9884777d11f70eb7900fffe9e5adf650fc", "versionedPdfMatchesArchivedBytes": true, "checkedAt": "2026-10-04T13:54:03.645251+00:00" } }, "originalAnalysis": { "id": "arxiv-2609.40245", "sectionsRead": [ "摘要，PDF第1页", "第3节方法，PDF第3–6页", "第4节实验与表1，PDF第6–9页", "第5–6节局限和结论，PDF第9页" ], "analyzedAt": "2026-10-04T13:53:31.184413+00:00", "analysisStatus": "full_text_sections", "methodsZh": "用完整有向图表示同时观察到的代理，边包含相对SE(2)位姿及时间差分。异构版本区分人、机器人与关系类型；消息传递图VAE重建节点有效属性和时序边特征，以KL正则学习每节点潜变量，再冻结编码器作线性探测。", "experimentsZh": "联合使用SEAN-T的八秒VR导航观察与SNS行人交互数据，评测能力、意图、惊讶等主观感知以及行人关系动作分类。逐步使用1%至100%标签，比较MLP、自编码器、随机森林及同构/异构图，十个随机种子。", "resultsZh": "SNS全标签宏F1为0.503，优于非结构化自编码器0.332；1%标签仍达0.362。SEAN-T上同构版本全标签能力/惊讶/意图F1为0.799/0.741/0.734，但部分低数据设置自编码器更好。", "limitationsZh": "来源警告：arXiv落地页摘要与同标识v2 PDF不一致。落地页摘要描述SocialNav-SUB视觉问答基准，v2 PDF则是STARS时空图自监督表征研究；已确认PDF标题、作者、标识和版本匹配。本站以v2 PDF分析，未沿用误配摘要；上游不一致原因尚未确定。 评测依赖处理好的轨迹和固定窗口，未验证在线分段、感知噪声或闭环导航。图聚合与压缩会平滑精细几何，对群体检测等边级任务可能不利；潜空间可视化只是定性证据。", "reproductionZh": "复现需固定图构建、有效节点掩码、标签比例、随机种子与冻结线性探测协议，避免把下游标签泄露到自监督阶段。本次以v2原文改正输入摘要错配，项目代码是否实际可下载未独立核验。", "sourceUrl": "https://arxiv.org/pdf/2609.40245v2", "sourceVersion": "2609.40245v2", "originalSha256": "9d87aa14858069890ac9bfcb61bfbe9884777d11f70eb7900fffe9e5adf650fc", "experimentType": "data", "robots": [ "SEAN-T虚拟导航代理", "SNS数据中的机器人与行人（离线轨迹）" ], "evidenceNotes": [ { "section": "PDF第1页摘要", "note": "arXiv v2（2026-10-01）是STARS图表征方法；输入abstractOriginal为SocialNav-SUB基准摘要，二者错配。" }, { "section": "PDF第8页表1（已渲染核看）", "note": "SNS宏F1=0.503；SEAN-T并非每个标注比例或指标都占优。" }, { "section": "PDF第7页实验与第9页局限", "note": "基于既有交互数据做离线预测，没有执行机器人导航政策。" } ], "verification": "partial", "directionsOriginal": [ "人机交互", "关系表征学习" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "metadataWarning": "来源警告：arXiv落地页摘要与同标识v2 PDF不一致。落地页摘要描述SocialNav-SUB视觉问答基准，v2 PDF则是STARS时空图自监督表征研究；已确认PDF标题、作者、标识和版本匹配。本站以v2 PDF分析，未沿用误配摘要；上游不一致原因尚未确定。" }, { "id": "arxiv-2609.40297", "title": "GPU-Accelerated Path-Dependent Marginal Information Gain for Autonomous Exploration", "titleZh": "用于自主探索的GPU加速路径相关边际信息增益计算", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2609.40297", "paperUrl": "https://arxiv.org/abs/2609.40297", "projectUrl": null, "codeUrl": null, "category": "导航与自主探索", "tags": [ "信息增益", "GPU加速", "无人机", "真机实验" ], "directions": [ "导航与建图" ], "robotFilters": [ "DJI F550" ], "tier": "recent", "summary": "以深度缓冲替代祖先视点体素集合去重，最高118倍计算加速；真机达到95%覆盖的时间约减少30%。", "abstractZh": "传统探索规划常独立计各视点信息增益，重复奖励沿路径可见的同一区域。论文把历史视点存为深度缓冲，将候选射线投影其中扣除预期已观察部分；按树深度顺序、同层并行计算以保持视角依赖。该近似插入现有采样规划器，计算与覆盖效率改善，但收益取决于环境和规划器。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "both", "experimentNote": "与精确CPU体素哈希法比较166440个节点，扫树大小和体素分辨率；RH-NBVP、AEP在学校、迷宫、多层环境各条件10次。Jetson作硬件在环计算测试；DJI F550搭载D455和Orin NX在10×13×6米庭院真实探索。", "robots": [ "DJI F550（真机）", "MRS UAV（Gazebo仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "应将GPU数据传输纳入计时，区分算法核加速与完整探索节省；保持两种增益的规划参数、偏航优化及初始状态一致。复现同时记录覆盖曲线、路径、速度与终止阈值，避免把更快运动误当更优探索。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "CPU构树与碰撞检测，将TSDF连续化后传GPU。每节点保存第一占据体素的深度图，候选射线在祖先图像中裁剪并以逆深度插值找重叠区间，再三维射线累积新增未知体素。逐深度处理节点及偏航，同层节点和射线并行。", "whyUseful": "应将GPU数据传输纳入计时，区分算法核加速与完整探索节省；保持两种增益的规划参数、偏航优化及初始状态一致。复现同时记录覆盖曲线、路径、速度与终止阈值，避免把更快运动误当更优探索。", "limitations": "近似不是精确信息增益，最坏时间阶仍与哈希法相同；绝对增益GPU计算仍更快。多层场景的RH-NBVP未获收益，采样可达性限制无法靠评分修复；仅一处真实庭院，不能外推任意动态环境。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2609.40297v1", "note": "PDF第6页表II（已渲染核看）：118倍发生于0.1 m/1000节点桌面配置；28倍为Jetson的0.1 m/500节点。" }, { "url": "https://arxiv.org/pdf/2609.40297v1", "note": "PDF第7页§VI：DJI F550真机闭环，95%覆盖2.17对3.10分钟，速度近似。" }, { "url": "https://arxiv.org/pdf/2609.40297v1", "note": "PDF第7页§V-C：六种组合之一RH-NBVP多层环境95%覆盖反而慢0.4分钟。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T14:02:45.217003+00:00", "original": { "id": "arxiv-2609.40297", "originalSourceUrl": "https://arxiv.org/pdf/2609.40297v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.40297v1", "finalDownloadUrl": "https://arxiv.org/pdf/2609.40297", "sha256": "6c77b23296c0777f5a6eb430ef3e790a5972def4c3e4c39ab4891d3e7aa9a9a5", "bytes": 1085579, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:44:46.862218+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.40297", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.40297", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.40297", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "GPU-Accelerated Path-Dependent Marginal Information Gain for Autonomous Exploration", "sourceVersion": "2609.40297v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.770682+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.40297" }, "originalAnalysis": { "id": "arxiv-2609.40297", "sectionsRead": [ "第III–IV节方法，PDF第2–5页", "第V节仿真实验、表I–IV，PDF第5–7页", "第VI节真机实验，PDF第7页", "第VII节结论开头，PDF第7页" ], "analyzedAt": "2026-10-04T14:02:45.217003+00:00", "analysisStatus": "full_text_sections", "methodsZh": "CPU构树与碰撞检测，将TSDF连续化后传GPU。每节点保存第一占据体素的深度图，候选射线在祖先图像中裁剪并以逆深度插值找重叠区间，再三维射线累积新增未知体素。逐深度处理节点及偏航，同层节点和射线并行。", "experimentsZh": "与精确CPU体素哈希法比较166440个节点，扫树大小和体素分辨率；RH-NBVP、AEP在学校、迷宫、多层环境各条件10次。Jetson作硬件在环计算测试；DJI F550搭载D455和Orin NX在10×13×6米庭院真实探索。", "resultsZh": "估计与精确值R²=0.9937，有约5–10%高估；桌面最高118倍、Jetson最高28倍加速。六种规划器—环境组合中五种降低95%覆盖时间。真机从3.10±1.05分钟降至2.17±0.44分钟，终止时间4.58降至4.04分钟。", "limitationsZh": "近似不是精确信息增益，最坏时间阶仍与哈希法相同；绝对增益GPU计算仍更快。多层场景的RH-NBVP未获收益，采样可达性限制无法靠评分修复；仅一处真实庭院，不能外推任意动态环境。", "reproductionZh": "应将GPU数据传输纳入计时，区分算法核加速与完整探索节省；保持两种增益的规划参数、偏航优化及初始状态一致。复现同时记录覆盖曲线、路径、速度与终止阈值，避免把更快运动误当更优探索。", "sourceUrl": "https://arxiv.org/pdf/2609.40297v1", "sourceVersion": "2609.40297v1", "originalSha256": "6c77b23296c0777f5a6eb430ef3e790a5972def4c3e4c39ab4891d3e7aa9a9a5", "experimentType": "both", "robots": [ "DJI F550（真机）", "MRS UAV（Gazebo仿真）" ], "evidenceNotes": [ { "section": "PDF第6页表II（已渲染核看）", "note": "118倍发生于0.1 m/1000节点桌面配置；28倍为Jetson的0.1 m/500节点。" }, { "section": "PDF第7页§VI", "note": "DJI F550真机闭环，95%覆盖2.17对3.10分钟，速度近似。" }, { "section": "PDF第7页§V-C", "note": "六种组合之一RH-NBVP多层环境95%覆盖反而慢0.4分钟。" } ], "verification": "verified", "directionsOriginal": [ "三维探索", "采样式运动规划" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Autonomous exploration demands that robots continuously evaluate candidate viewpoints based on their expected information gain and execution cost. Sampling-based planners estimate this gain by volumetric raycasting and, due to its computational cost, evaluate candidates under an assumption of mutual independence, ignoring the overlap between viewpoints along the same path. This work presents a GPU-accelerated method for computing path-dependent marginal information gain, where instead of storing and merging the observed unknown voxels along each candidate path, previous observations are represented using depth buffers. Candidate rays are projected into the depth buffers of their ancestors to identify observation overlap and exclude regions expected to be observed. The planning tree is evaluated in depth order to maintain the dependency between viewpoints and their optimized yaws, while candidate nodes and rays at each level are processed in parallel on the GPU. The proposed method stays within 5-10% of the exact marginal gain computed using voxel hash maps, with speed-ups of up to 118x on a desktop GPU and 28x on an NVIDIA Jetson Orin NX. The method was integrated into two sampling-based exploration planners and evaluated in three simulation environments, where marginal gain reduced the time to 95% coverage in five of the six evaluated planner-environment combinations. Real-world experiments also showed a 30% reduction in the time to 95% coverage, as well as earlier exploration termination times." }, { "id": "arxiv-2609.40306", "title": "DynaHarness: A Dynamic Physical Harness for Self-Evolving Robot Agents", "titleZh": "DynaHarness：用于自演化机器人智能体的动态物理执行框架", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2609.40306", "paperUrl": "https://arxiv.org/abs/2609.40306", "projectUrl": null, "codeUrl": null, "category": "机器人智能体与执行控制", "tags": [ "运行时执行", "技能库", "失败归因", "真机实验" ], "directions": [ "视觉语言动作" ], "robotFilters": [ "Universal Robots UR7e" ], "tier": "recent", "summary": "以执行约束、技能替换与回归验证组织机器人智能体，新初始状态成功率75.2%；主要能力来自解析技能而非VLA本体进化。", "abstractZh": "DynaHarness让慢速语言规划器提出符号技能，快速确定性执行层负责几何落地、拒绝不可执行指令、限定预算、监测与改道；所有行动共享可追溯执行记录。离线归因指引可复用技能修订，再用配对和广泛回归决定是否采用。它提升的是整套系统的技能与控制组织，冻结VLA权重本身并未更新。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "both", "experimentNote": "主实验为LIBERO-Pro四单元共800开发回合，冻结后再测800新状态及261未触及官方状态。同技能库对照动态执行与普通逐步重规划。UR7e搭配双RealSense和SAM3完成四项真机任务，各10次位置变化试验。", "robots": [ "UR7e（真机）", "LIBERO机械臂（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现要冻结最终配置再采样测试，分别保存开发、重测、官方与新状态结果；保留解析技能、来源状态及成功检测频率。不能把75.2%归因于模型微调，也不能把运行时50 Hz安全检查写成可证明真实安全。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "慢脑Qwen3-VL-4B按需输出技能和符号参数；2 Hz快脑以几何条件、预算和租期决定接纳、拒绝、替换或重规划，20 Hz执行并锁存短暂成功事件。解析、恢复技能及冻结π0.5统一记录，失败按有序层级定位，候选修订经配对和全套回归筛选。", "whyUseful": "复现要冻结最终配置再采样测试，分别保存开发、重测、官方与新状态结果；保留解析技能、来源状态及成功检测频率。不能把75.2%归因于模型微调，也不能把运行时50 Hz安全检查写成可证明真实安全。", "limitations": "仿真几何参数来自模拟器状态，完成信号取基准成功谓词；新状态迁移仍是相同任务单元。LIBERO-Plus每任务仅一试且无冻结政策对照。失败归因是诊断假设，修订过程自动化范围不能扩大为无人介入持续自学习。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2609.40306v1", "note": "PDF第7页表2–3（已渲染核看）：75.2%为新状态；真机明确四任务各10次，非仅仿真。" }, { "url": "https://arxiv.org/pdf/2609.40306v1", "note": "附录A，PDF第14页附近：LIBERO能力落地几何来自simulator state；硬件场景来自相机。" }, { "url": "https://arxiv.org/pdf/2609.40306v1", "note": "PDF第8–9页§4.5、4.7：770归档回合中570次未调用VLA；去解析技能成功率16.6%。" }, { "url": "https://arxiv.org/pdf/2609.40306v1", "note": "PDF第37页附录I：自动化覆盖归因和验证；任务与物理域泛化仍须分别测。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T14:02:45.215685+00:00", "original": { "id": "arxiv-2609.40306", "originalSourceUrl": "https://arxiv.org/pdf/2609.40306v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.40306v1", "finalDownloadUrl": "https://arxiv.org/pdf/2609.40306", "sha256": "e07c7fd79a2e6690f64cbc07f501894bc96d07c9bc091a735a7ef403572ddd1d", "bytes": 7704124, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:44:40.225343+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.40306", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.40306", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.40306", "pages": 37, "sourceIdentityStatus": "title_match", "sourceTitle": "DynaHarness: A Dynamic Physical Harness for Self-Evolving Robot Agents", "sourceVersion": "2609.40306v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.768148+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.40306" }, "originalAnalysis": { "id": "arxiv-2609.40306", "sectionsRead": [ "第3节方法，PDF第3–5页", "第4节实验、表1–3，PDF第5–10页", "附录A的执行接口与状态来源，PDF第14–15页（选段）", "附录B.1真机平台与协议，PDF第18页", "附录I局限，PDF第37页" ], "analyzedAt": "2026-10-04T14:02:45.215685+00:00", "analysisStatus": "full_text_sections", "methodsZh": "慢脑Qwen3-VL-4B按需输出技能和符号参数；2 Hz快脑以几何条件、预算和租期决定接纳、拒绝、替换或重规划，20 Hz执行并锁存短暂成功事件。解析、恢复技能及冻结π0.5统一记录，失败按有序层级定位，候选修订经配对和全套回归筛选。", "experimentsZh": "主实验为LIBERO-Pro四单元共800开发回合，冻结后再测800新状态及261未触及官方状态。同技能库对照动态执行与普通逐步重规划。UR7e搭配双RealSense和SAM3完成四项真机任务，各10次位置变化试验。", "resultsZh": "新状态成功率75.2%，冻结π0.5为17.5%；同库动态执行74.0%，普通重规划63.9%。去除解析接触技能降至16.6%，说明优势主要由技能库供给。真机四任务90%、80%、70%、70%；部分局部改进通过小检验却在全套回归退化而被拒绝。", "limitationsZh": "仿真几何参数来自模拟器状态，完成信号取基准成功谓词；新状态迁移仍是相同任务单元。LIBERO-Plus每任务仅一试且无冻结政策对照。失败归因是诊断假设，修订过程自动化范围不能扩大为无人介入持续自学习。", "reproductionZh": "复现要冻结最终配置再采样测试，分别保存开发、重测、官方与新状态结果；保留解析技能、来源状态及成功检测频率。不能把75.2%归因于模型微调，也不能把运行时50 Hz安全检查写成可证明真实安全。", "sourceUrl": "https://arxiv.org/pdf/2609.40306v1", "sourceVersion": "2609.40306v1", "originalSha256": "e07c7fd79a2e6690f64cbc07f501894bc96d07c9bc091a735a7ef403572ddd1d", "experimentType": "both", "robots": [ "UR7e（真机）", "LIBERO机械臂（仿真）" ], "evidenceNotes": [ { "section": "PDF第7页表2–3（已渲染核看）", "note": "75.2%为新状态；真机明确四任务各10次，非仅仿真。" }, { "section": "附录A，PDF第14页附近", "note": "LIBERO能力落地几何来自simulator state；硬件场景来自相机。" }, { "section": "PDF第8–9页§4.5、4.7", "note": "770归档回合中570次未调用VLA；去解析技能成功率16.6%。" }, { "section": "PDF第37页附录I", "note": "自动化覆盖归因和验证；任务与物理域泛化仍须分别测。" } ], "verification": "verified", "directionsOriginal": [ "分层机器人智能体", "验证式系统改进" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Pretrained robot policies provide useful action priors, but long-horizon manipulation still requires coordination between semantic reasoning and physical execution. Semantic reasoning operates at a coarser timescale than physical interaction, while episode-level failures provide limited guidance on which system component should be revised. We propose DynaHarness, a dynamic physical harness that couples semantic reasoning with physical governance through a shared execution contract and turns failure evidence into validated capability revisions. To be more specific, the slow brain proposes capabilities and symbolic arguments, while the fast brain grounds and monitors commands, refuses unresolved actions, substitutes capabilities, and requests replans when needed. The physical execution contract bounds each accepted command and records execution evidence across analytic skills, recovery skills, and the frozen VLA. Failure attribution localizes faults in these records and directs targeted revisions of reusable capabilities or execution mechanisms. Paired regression checks govern admission or rejection, closing the self-evolution loop. On LIBERO-Pro, DynaHarness achieves 75.2% on 800 newly sampled initial states, compared with 17.5% for the frozen policy. With the same capability library, full dynamic execution reaches 74.0% versus 63.9% under nominal one-step replanning. This demonstrates the value of DynaHarness as a dynamic physical harness that governs how existing capabilities are grounded, monitored, and coordinated during execution. Our project page is at https://denghaoyuan123.github.io/Dynaharness_page/." }, { "id": "arxiv-2609.40341", "title": "Ego4WAM: What Matters When Scaling Egocentric Human Data for Robot Learning?", "titleZh": "Ego4WAM：扩大第一人称人类数据规模用于机器人学习时，什么最重要？", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2609.40341", "paperUrl": "https://arxiv.org/abs/2609.40341", "projectUrl": null, "codeUrl": null, "category": "人类数据与机器人学习", "tags": [ "第一人称视频", "数据规模", "世界动作模型", "双臂操作" ], "directions": [ "世界模型", "模仿学习" ], "robotFilters": [ "AgileX PiPER" ], "tier": "recent", "summary": "受控实验表明数据时长、任务多样性和使用阶段作用不同；视频预训练加未来视觉条件可显著改善机器人操作。", "abstractZh": "Ego4WAM在固定世界动作模型框架内研究第一人称数据的对齐程度、动作标签、时长和多样性。少量任务对齐的人类示范帮助新对象与场景迁移，广泛多任务经验提高少样本适应；无可靠动作标注的视频仍可学习动态先验。规模收益并非单调，人类数据也不能完全替代具身机器人示范。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "both", "experimentNote": "双臂Piper真机测试装篮与叠布：每任务300条机器人示范，500条含对象/场景变化的人类示范。RoboDojo仿真固定常见500任务扩时长，或扩至6000任务；视频预训练15000小时，可靠动作中间训练12000小时并含约90小时机器人数据。", "robots": [ "双臂Piper（真机）", "RoboDojo仿真机器人" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现需固定数据质量过滤、各来源比例、动作缺失掩码和未来注意可见性，并拆分任务数与每任务时长。论文组织的数据池约12万小时纯视频，不等于本实验实际训练用了12万小时。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "把数据分成任务对齐视频动作、多任务可靠动作与纯视频三层。世界视频分支和动作分支共享注意，双臂32维动作只在人机共同可解释维度监督；依次尝试视频预训练、视频动作中间训练与机器人适配，并比较动作能否读取预测未来视觉。", "whyUseful": "复现需固定数据质量过滤、各来源比例、动作缺失掩码和未来注意可见性，并拆分任务数与每任务时长。论文组织的数据池约12万小时纯视频，不等于本实验实际训练用了12万小时。", "limitations": "纯视频与带动作阶段不等计算预算，不能只归因为监督类型。最佳配置在开放指令与记忆类仍弱，14.35%并非榜单最佳；人类侧见过的分布外对象不是完全未见对象。额外四项真机展示未给同等完整统计。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2609.40341v1", "note": "PDF第8页表1–2（已渲染核看）：最终14.35%配置是video pretrain + joint posttrain，表中没有mid-train；不能笼统写三阶段全部叠加最佳。" }, { "url": "https://arxiv.org/pdf/2609.40341v1", "note": "PDF第7–8页§4.3：视频预训练实际为15K小时，video-action为12K小时，明确非compute-matched。" }, { "url": "https://arxiv.org/pdf/2609.40341v1", "note": "PDF第5–6页图3–4：真正物理执行为双臂Piper；对象OOD由人类演示覆盖。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T14:02:45.214874+00:00", "original": { "id": "arxiv-2609.40341", "originalSourceUrl": "https://arxiv.org/pdf/2609.40341v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.40341v1", "finalDownloadUrl": "https://arxiv.org/pdf/2609.40341", "sha256": "09789e8fe4711a3c3fdd578eebd8e681ca91a9e36d525346d7dd440de3fdfaab", "bytes": 7262487, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:44:34.674497+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.40341", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.40341", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.40341", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "Ego4WAM: What Matters When Scaling Egocentric Human Data for Robot Learning?", "sourceVersion": "2609.40341v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.760769+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.40341" }, "originalAnalysis": { "id": "arxiv-2609.40341", "sectionsRead": [ "第3节方法，PDF第3–5页", "第4节实验及表1–2，PDF第5–9页", "第5节结论，PDF第9页", "附录C.1–C.2设置，PDF第16页" ], "analyzedAt": "2026-10-04T14:02:45.214874+00:00", "analysisStatus": "full_text_sections", "methodsZh": "把数据分成任务对齐视频动作、多任务可靠动作与纯视频三层。世界视频分支和动作分支共享注意，双臂32维动作只在人机共同可解释维度监督；依次尝试视频预训练、视频动作中间训练与机器人适配，并比较动作能否读取预测未来视觉。", "experimentsZh": "双臂Piper真机测试装篮与叠布：每任务300条机器人示范，500条含对象/场景变化的人类示范。RoboDojo仿真固定常见500任务扩时长，或扩至6000任务；视频预训练15000小时，可靠动作中间训练12000小时并含约90小时机器人数据。", "resultsZh": "装篮对象分布外成功率可从10%升至60%。仅20条叠布机器人示范加对齐人类数据仍失败，加入多任务中间训练后ID达80%。RoboDojo基线成功率3.15%，视频预训练9.45%，未来视觉联合条件14.35%；长尾扩展可能损害精密操作。", "limitationsZh": "纯视频与带动作阶段不等计算预算，不能只归因为监督类型。最佳配置在开放指令与记忆类仍弱，14.35%并非榜单最佳；人类侧见过的分布外对象不是完全未见对象。额外四项真机展示未给同等完整统计。", "reproductionZh": "复现需固定数据质量过滤、各来源比例、动作缺失掩码和未来注意可见性，并拆分任务数与每任务时长。论文组织的数据池约12万小时纯视频，不等于本实验实际训练用了12万小时。", "sourceUrl": "https://arxiv.org/pdf/2609.40341v1", "sourceVersion": "2609.40341v1", "originalSha256": "09789e8fe4711a3c3fdd578eebd8e681ca91a9e36d525346d7dd440de3fdfaab", "experimentType": "both", "robots": [ "双臂Piper（真机）", "RoboDojo仿真机器人" ], "evidenceNotes": [ { "section": "PDF第8页表1–2（已渲染核看）", "note": "最终14.35%配置是video pretrain + joint posttrain，表中没有mid-train；不能笼统写三阶段全部叠加最佳。" }, { "section": "PDF第7–8页§4.3", "note": "视频预训练实际为15K小时，video-action为12K小时，明确非compute-matched。" }, { "section": "PDF第5–6页图3–4", "note": "真正物理执行为双臂Piper；对象OOD由人类演示覆盖。" } ], "verification": "verified", "directionsOriginal": [ "人机跨具身迁移", "机器人预训练" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Egocentric human data provides a scalable source of experience for robot learning, but varies substantially in human-robot alignment, behavioral coverage, and available supervision. Existing work shows favorable scaling with increasing human data, but it remains unclear which data properties drive downstream robot gains and how to use such data throughout the training pipeline. We present a systematic study of egocentric human data with different alignment and supervision under a unified world-action model framework. With the model backbone fixed, we disentangle the effects of human-robot alignment, data duration and task diversity, action supervision, and data usage strategies. We find that aligned human demonstrations substantially improve out-of-distribution generalization and reduce target-task robot data requirements; data duration and task diversity affect downstream capabilities differently; and video-only supervision remains effective without action labels, providing a strong foundation for subsequent video-action training. We validate these findings through closed-loop policy evaluation on both real robots and RoboDojo. Rather than treating data duration as the sole scaling axis, Ego4WAM shows how alignment, task diversity, available supervision, and usage strategy jointly shape the value of egocentric human data for robot learning." }, { "id": "arxiv-2609.40353", "title": "AssemblyWorld: Rethinking 3D Assembly with General-Purpose Agents", "titleZh": "AssemblyWorld：用通用智能体重新思考三维装配", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2609.40353", "paperUrl": "https://arxiv.org/abs/2609.40353", "projectUrl": null, "codeUrl": null, "category": "评测与数据集", "tags": [ "工具使用", "三维几何", "装配基准", "交互式智能体" ], "directions": [ "数据集与基准" ], "robotFilters": [], "tier": "recent", "summary": "在统一交互式几何环境中比较八种智能体，最佳整体装配成功率59.4%；评测不包含物理碰撞或稳定装配执行。", "abstractZh": "AssemblyWorld通过MCP提供视角观察与部件位姿编辑，让通用模型在不做装配专用微调的前提下尝试家具、工业零件及碎片重组。代理看渲染图和受限状态查询，不能直接读网格顶点。基准按最终几何评分，揭示通用推理、接口可靠性和精细位姿之间的差异，不能直接等同于机器人完成实体装配。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "从PartNet、IKEA-Manual、AssemblyBench、Fantastic Breaks各取20对象，PartNet有/无参考图形成100任务。八种模型—执行框架组合共享工具和60分钟上限，另以更大来源子集对照专用方法并测试GARF精修顺序。", "robots": [ "无实体机器人；MCP三维部件编辑环境" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "需保存模型版本、框架、超时与错误终止记录，固定布局及来源权重，区分SR与用整体形状距离定义的SRa。专用方法采用各自报告协议，不能将全部表格当成同输入、同预算的严格控制实验。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "固定部件几何和尺度，随机偏航与散布初始化，通过渲染、单件/组平移旋转工具逐步组合；网格顶点与面不开放。最终全局对齐并用匈牙利匹配处理等价部件，以Chamfer距离、部件准确率及全件正确成功率评估。", "whyUseful": "需保存模型版本、框架、超时与错误终止记录，固定布局及来源权重，区分SR与用整体形状距离定义的SRa。专用方法采用各自报告协议，不能将全部表格当成同输入、同预算的严格控制实验。", "limitations": "所有结论针对自由空间几何编辑，无碰撞、稳定性、抓取或可达性验证。整体分数是数据源等权平均，不是100题简单平均；模型与执行框架混合比较，预训练暴露未知，小样本单次鲁棒性结论有限。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2609.40353v1", "note": "PDF第5页表1（已渲染核看）：overall先平均PartNet两种条件，再对四个数据源等权；59.4%不是未经加权的任务计数。" }, { "url": "https://arxiv.org/pdf/2609.40353v1", "note": "PDF第9页结论：明确只研究free-space geometry，不建立collision-free或physically stable assembly。" }, { "url": "https://arxiv.org/pdf/2609.40353v1", "note": "PDF第9页表6：GARF→Agent PA=93.00%，Agent→GARF=87.00%；精修并非任意次序都有效。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:57:30.178020+00:00", "original": { "id": "arxiv-2609.40353", "originalSourceUrl": "https://arxiv.org/pdf/2609.40353v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.40353v1", "finalDownloadUrl": "https://arxiv.org/pdf/2609.40353", "sha256": "97d311f41e51e9e96c1176176faad8f4ac518691794b3581a2a5703e121cb289", "bytes": 3386510, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:44:34.560316+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.40353", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.40353", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.40353", "pages": 24, "sourceIdentityStatus": "title_match", "sourceTitle": "AssemblyWorld: Rethinking 3D Assembly with General-Purpose Agents", "sourceVersion": "2609.40353v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.753031+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.40353" }, "originalAnalysis": { "id": "arxiv-2609.40353", "sectionsRead": [ "第3节环境与基准，PDF第3–4页", "第4节实验及表1–6，PDF第4–9页", "第5节结论局限，PDF第9页" ], "analyzedAt": "2026-10-04T13:57:30.178020+00:00", "analysisStatus": "full_text_sections", "methodsZh": "固定部件几何和尺度，随机偏航与散布初始化，通过渲染、单件/组平移旋转工具逐步组合；网格顶点与面不开放。最终全局对齐并用匈牙利匹配处理等价部件，以Chamfer距离、部件准确率及全件正确成功率评估。", "experimentsZh": "从PartNet、IKEA-Manual、AssemblyBench、Fantastic Breaks各取20对象，PartNet有/无参考图形成100任务。八种模型—执行框架组合共享工具和60分钟上限，另以更大来源子集对照专用方法并测试GARF精修顺序。", "resultsZh": "GPT-6 Astra整体成功率59.4%，Fable为50.0%，Opus为44.4%；Astra平均5.8分钟、94次工具调用。两碎片任务可到90%，图像条件PartNet仅40%。GARF后由智能体修正得93.00%部件准确率，反向组合反而可能退化。", "limitationsZh": "所有结论针对自由空间几何编辑，无碰撞、稳定性、抓取或可达性验证。整体分数是数据源等权平均，不是100题简单平均；模型与执行框架混合比较，预训练暴露未知，小样本单次鲁棒性结论有限。", "reproductionZh": "需保存模型版本、框架、超时与错误终止记录，固定布局及来源权重，区分SR与用整体形状距离定义的SRa。专用方法采用各自报告协议，不能将全部表格当成同输入、同预算的严格控制实验。", "sourceUrl": "https://arxiv.org/pdf/2609.40353v1", "sourceVersion": "2609.40353v1", "originalSha256": "97d311f41e51e9e96c1176176faad8f4ac518691794b3581a2a5703e121cb289", "experimentType": "sim", "robots": [ "无实体机器人；MCP三维部件编辑环境" ], "evidenceNotes": [ { "section": "PDF第5页表1（已渲染核看）", "note": "overall先平均PartNet两种条件，再对四个数据源等权；59.4%不是未经加权的任务计数。" }, { "section": "PDF第9页结论", "note": "明确只研究free-space geometry，不建立collision-free或physically stable assembly。" }, { "section": "PDF第9页表6", "note": "GARF→Agent PA=93.00%，Agent→GARF=87.00%；精修并非任意次序都有效。" } ], "verification": "verified", "directionsOriginal": [ "几何推理", "通用智能体评测" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "The task of 3D assembly requires translating an understanding of parts and their relationships into precise spatial arrangements. Can pretrained general-purpose agents assemble objects through visual interaction without additional assembly-specific fine-tuning? To investigate this question, we introduce AssemblyWorld, an interactive 3D environment in which agents inspect rendered views and manipulate supplied rigid parts, guided by images or assembly manuals when available. Agents perceive part geometry through 2D views rather than direct access to mesh vertices or faces, while their resulting assemblies are evaluated geometrically. Building on this environment, we construct AssemblyWorldBench, comprising 100 assembly tasks across 80 objects spanning furniture, industrial assembly, and fracture reassembly. Evaluating eight agent systems reveals substantial differences in their capabilities. The strongest system achieves 80.9% part accuracy but 59.4% complete-assembly success. The evaluated open-source systems lag substantially behind their stronger closed-source peers in both execution reliability and assembly accuracy. Analyses of visual references, interaction trajectories, and failures show how agents revise assemblies while leaving residual positioning errors. AssemblyWorld provides a common setting for both assessing the capabilities of interactive assembly agents and characterizing the gap between approximate structure recovery and precise reconstruction." }, { "id": "arxiv-2610.00438", "title": "Towards a General Humanoid Loco-Manipulation Model via Egocentric Whole-Body Human Data Pretraining", "titleZh": "通过第一人称全身人类数据预训练迈向通用人形机器人运动操作模型", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00438", "paperUrl": "https://arxiv.org/abs/2610.00438", "projectUrl": null, "codeUrl": null, "category": "人形机器人与全身操作", "tags": [ "全身控制", "人类数据", "第一人称", "运动操作" ], "directions": [ "视觉语言动作", "运动控制", "模仿学习" ], "robotFilters": [ "Unitree G1", "BrainCo Revo 2" ], "tier": "recent", "summary": "HumanVerse-500配合三阶段λ0训练提升全身协调，SIMPLE成功率90.0%，G1四项真机任务25/40成功。", "abstractZh": "论文发布500小时同步第一人称视频与全身、双手动作的数据方案，并训练共享视觉语言动作主干λ0。先学习多源人类手物交互，再从HumanVerse学习身体与手的协调，最后用机器人示范适配。结果支持全身人类中间训练的价值，但只验证Unitree G1，机器人未见对象也可能已在人类数据中出现。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "both", "experimentNote": "三阶段依次使用七个人类手物数据源、500小时全身数据和机器人遥操作。SIMPLE六任务三级各10回合，真实G1配Revo2双手在四任务各10个固定初始配置测试；消融人类训练阶段并扫描5%至100%全身数据。", "robots": [ "Unitree G1 + BrainCo Revo 2双手（真机）", "SIMPLE人形机器人（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "应保留动作域标识、有效性掩码、64维SONIC与双6维手命令接口，并记录不同阶段EMA初始化。比较数据规模要匹配机器人数据、更新预算和检查点选择，明确human-guided不同于完全零样本。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "HumanVerse-500含32993回合、827任务类型及8700万姿态帧，PICO追踪身体、视频恢复手部并做时空对齐。λ0用Qwen3.5-2B及流匹配动作专家共享迁移核心，域专用投影区分人机；SONIC把人体动作重定向为G1运动潜变量并控制全身。", "whyUseful": "应保留动作域标识、有效性掩码、64维SONIC与双6维手命令接口，并记录不同阶段EMA初始化。比较数据规模要匹配机器人数据、更新预算和检查点选择，明确human-guided不同于完全零样本。", "limitations": "只验证G1，跨大型人形形态未测；重定向、相机高度与身体比例仍有偏差。每任务仅10次，规模拟合为单次运行且仅描述测量区间；部分仿真基线保留来源协议，不能视作全部同预算重跑。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00438v1", "note": "PDF第8页表1–2（已渲染核看）：真实25/40、仿真90.0%；进度与成功是不同指标，椅子任务进度并非最高。" }, { "url": "https://arxiv.org/pdf/2610.00438v1", "note": "PDF第9页表3：去Stage II为22.5%真机成功，去Stage I为50.0%，两个阶段贡献不同。" }, { "url": "https://arxiv.org/pdf/2610.00438v1", "note": "PDF第9–10页§5.4及图10：Human-guided对象B只在人的演示出现，不能宣称所有训练均未见。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T14:02:45.218640+00:00", "original": { "id": "arxiv-2610.00438", "originalSourceUrl": "https://arxiv.org/pdf/2610.00438v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00438v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00438", "sha256": "6d2e48b284c9d2d6a21882239cb6dcec5b493744c1bac0bcd2681008849fbc76", "bytes": 39808001, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:45:01.091998+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00438", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00438", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00438", "pages": 15, "sourceIdentityStatus": "title_match", "sourceTitle": "Towards a General Humanoid Loco-Manipulation Model via Egocentric Whole-Body Human Data Pretraining", "sourceVersion": "2610.00438v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.824771+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00438" }, "originalAnalysis": { "id": "arxiv-2610.00438", "sectionsRead": [ "第3节数据采集与重定向，PDF第3–5页", "第4节策略与训练，PDF第5–6页", "第5节实验、表1–3及图8–10，PDF第7–10页", "第6节局限，PDF第10页" ], "analyzedAt": "2026-10-04T14:02:45.218640+00:00", "analysisStatus": "full_text_sections", "methodsZh": "HumanVerse-500含32993回合、827任务类型及8700万姿态帧，PICO追踪身体、视频恢复手部并做时空对齐。λ0用Qwen3.5-2B及流匹配动作专家共享迁移核心，域专用投影区分人机；SONIC把人体动作重定向为G1运动潜变量并控制全身。", "experimentsZh": "三阶段依次使用七个人类手物数据源、500小时全身数据和机器人遥操作。SIMPLE六任务三级各10回合，真实G1配Revo2双手在四任务各10个固定初始配置测试；消融人类训练阶段并扫描5%至100%全身数据。", "resultsZh": "仿真整体成功率90.0%，高于StarVLA86.1%；真机25/40即62.5%，进度80.4%，π0.5为40.0%成功。去掉全身中间训练真机降至22.5%；机器人未见但人类数据见过对象的平均进度55.8%，机器人单独训练零样本为48.4%。", "limitationsZh": "只验证G1，跨大型人形形态未测；重定向、相机高度与身体比例仍有偏差。每任务仅10次，规模拟合为单次运行且仅描述测量区间；部分仿真基线保留来源协议，不能视作全部同预算重跑。", "reproductionZh": "应保留动作域标识、有效性掩码、64维SONIC与双6维手命令接口，并记录不同阶段EMA初始化。比较数据规模要匹配机器人数据、更新预算和检查点选择，明确human-guided不同于完全零样本。", "sourceUrl": "https://arxiv.org/pdf/2610.00438v1", "sourceVersion": "2610.00438v1", "originalSha256": "6d2e48b284c9d2d6a21882239cb6dcec5b493744c1bac0bcd2681008849fbc76", "experimentType": "both", "robots": [ "Unitree G1 + BrainCo Revo 2双手（真机）", "SIMPLE人形机器人（仿真）" ], "evidenceNotes": [ { "section": "PDF第8页表1–2（已渲染核看）", "note": "真实25/40、仿真90.0%；进度与成功是不同指标，椅子任务进度并非最高。" }, { "section": "PDF第9页表3", "note": "去Stage II为22.5%真机成功，去Stage I为50.0%，两个阶段贡献不同。" }, { "section": "PDF第9–10页§5.4及图10", "note": "Human-guided对象B只在人的演示出现，不能宣称所有训练均未见。" } ], "verification": "verified", "directionsOriginal": [ "人形视觉语言动作模型", "跨具身预训练" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Humanoid whole-body manipulation has advanced rapidly, enabling policies to coordinate locomotion, posture, bimanual interaction, and dexterous hand movements. Meanwhile, egocentric human videos provide diverse examples of everyday interactions across objects and scenes, offering scalable supervision without robot operation. However, existing supervision from these videos provides limited coverage of whole-body movement and coordination with hand-object interaction, while obtaining such supervision through humanoid teleoperation is also costly and difficult to scale. We therefore explore how human experience can support scalable learning of humanoid loco-manipulation. To support this study, we introduce HumanVerse-500, a 500-hour dataset of diverse human loco-manipulation behaviors in open-world environments, collected with a lightweight wearable system that synchronizes egocentric video with body and hand motion. Building on this dataset, we develop $λ_0$, a whole-body humanoid vision-language-action policy, through three-stage training that first learns interaction from diverse egocentric datasets, then coordinates body and hand motion using HumanVerse-500, and finally adapts the policy to downstream tasks and robot embodiments. Across these stages, $λ_0$ learns a shared representation space for human experience transfer, while domain-specific interfaces handle differences between human and robot states and actions. We evaluate $λ_0$ on SIMPLE and 4 real-world loco-manipulation tasks, achieving state-of-the-art performance, and further analyze its scaling behavior, generalization, and training-stage contributions to understand how human data support downstream whole-body humanoid control. We will release our code, models, and data to support further research." }, { "id": "arxiv-2610.00487", "title": "ScaffoldM3C: A Multimodal Sequential Monte Carlo Framework for Generative Stable Construction Planning", "titleZh": "ScaffoldM3C：面向生成式稳定建造规划的多模态序贯蒙特卡洛框架", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00487", "paperUrl": "https://arxiv.org/abs/2610.00487", "projectUrl": null, "codeUrl": null, "category": "机器人装配与建造", "tags": [ "装配规划", "脚手架", "序贯蒙特卡洛", "真机示范" ], "directions": [], "robotFilters": [], "tier": "recent", "summary": "通过多候选积木生成与支撑块规划加快建造搜索；100%无碰撞不等于100%稳定，文本测试完整可行率为26.72%。", "abstractZh": "ScaffoldM3C把建造视为多模态条件下的下一块分布预测，引入可调高的支撑块词元，并用序贯蒙特卡洛维护多条装配假设。数据由既有序列补充支撑与多模态条件生成。实验显示相较BrickGPT可更快地产生无交叠结构，但在线搜索没有完整稳定性检查，实际硬件示范另加入拒绝采样筛选。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "both", "experimentNote": "仿真评测479条文本和288条图像提示，对照BrickGPT有/无拒绝采样及支撑版本；24条提示做粒子数和温度消融。六自由度机械臂实际搭建船、车、书架三种结构，使用45、86、89块，硬件计划额外经拒绝筛选。", "robots": [ "六自由度xArm Lite/XFactory机械臂（原文命名待统一）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "应区分无回滚的主表协议和加拒绝采样的硬件流程，固定候选/类别温度及粒子数，分别报告逐块和整结构稳定性。硬件名称在文中有xArm Lite与XFactory两种表述，具体型号需另核实。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "225M自回归Transformer结合Gemma条件编码，级联预测候选位置、形状与概率。SMC并行保留装配序列，二维最小平移修复碰撞、惩罚大位移并重采样，最终优先选最长序列；数据按重心投影支撑条件插入脚手架。", "whyUseful": "应区分无回滚的主表协议和加拒绝采样的硬件流程，固定候选/类别温度及粒子数，分别报告逐块和整结构稳定性。硬件名称在文中有xArm Lite与XFactory两种表述，具体型号需另核实。", "limitations": "搜索主要靠学到的稳定先验，不能保证所有中间状态稳定；最长序列选择偏向较大结构，未显式处理机械臂可达性。提示歧义、错位、漏块与早停仍存在，三次类型示范并非统计成功率评测。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00487v1", "note": "PDF第7页表I（已渲染核看）：100%是collision-free；文本overall stability/feasibility为26.72%，不能宣称稳定性保证。" }, { "url": "https://arxiv.org/pdf/2610.00487v1", "note": "PDF第8页§VI-E：真机前把SMC与rejection sampling结合，并把可变高支撑词元实例化为叠放单块。" }, { "url": "https://arxiv.org/pdf/2610.00487v1", "note": "PDF第6页与第8页：平台分别称xArm Lite和6-DOF XFactory，保存原文命名差异。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:57:30.177501+00:00", "original": { "id": "arxiv-2610.00487", "originalSourceUrl": "https://arxiv.org/pdf/2610.00487v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00487v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00487", "sha256": "28d7c3778489a86cc7e23cfe0cc75b2cbd8269b0684183cf747d15ea8669dd90", "bytes": 8615629, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:44:29.542916+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00487", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00487", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00487", "pages": 13, "sourceIdentityStatus": "title_match", "sourceTitle": "ScaffoldM3C: A Multimodal Sequential Monte Carlo Framework for Generative Stable Construction Planning", "sourceVersion": "2610.00487v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.749043+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00487" }, "originalAnalysis": { "id": "arxiv-2610.00487", "sectionsRead": [ "第III–V节方法及数据，PDF第3–5页", "第VI节实验与表I、图6，PDF第5–8页", "第VII–VIII节消融与局限，PDF第9–10页" ], "analyzedAt": "2026-10-04T13:57:30.177501+00:00", "analysisStatus": "full_text_sections", "methodsZh": "225M自回归Transformer结合Gemma条件编码，级联预测候选位置、形状与概率。SMC并行保留装配序列，二维最小平移修复碰撞、惩罚大位移并重采样，最终优先选最长序列；数据按重心投影支撑条件插入脚手架。", "experimentsZh": "仿真评测479条文本和288条图像提示，对照BrickGPT有/无拒绝采样及支撑版本；24条提示做粒子数和温度消融。六自由度机械臂实际搭建船、车、书架三种结构，使用45、86、89块，硬件计划额外经拒绝筛选。", "resultsZh": "文本SMC无碰撞率100%，完整稳定且无碰撞率26.72%，支撑版BrickGPT为11.27%。文本平均生成1.79秒，对照64.51秒；图像完整可行率20.49%。结构语义相似度接近基线，更多粒子提高质量但计算更贵。", "limitationsZh": "搜索主要靠学到的稳定先验，不能保证所有中间状态稳定；最长序列选择偏向较大结构，未显式处理机械臂可达性。提示歧义、错位、漏块与早停仍存在，三次类型示范并非统计成功率评测。", "reproductionZh": "应区分无回滚的主表协议和加拒绝采样的硬件流程，固定候选/类别温度及粒子数，分别报告逐块和整结构稳定性。硬件名称在文中有xArm Lite与XFactory两种表述，具体型号需另核实。", "sourceUrl": "https://arxiv.org/pdf/2610.00487v1", "sourceVersion": "2610.00487v1", "originalSha256": "28d7c3778489a86cc7e23cfe0cc75b2cbd8269b0684183cf747d15ea8669dd90", "experimentType": "both", "robots": [ "六自由度xArm Lite/XFactory机械臂（原文命名待统一）" ], "evidenceNotes": [ { "section": "PDF第7页表I（已渲染核看）", "note": "100%是collision-free；文本overall stability/feasibility为26.72%，不能宣称稳定性保证。" }, { "section": "PDF第8页§VI-E", "note": "真机前把SMC与rejection sampling结合，并把可变高支撑词元实例化为叠放单块。" }, { "section": "PDF第6页与第8页", "note": "平台分别称xArm Lite和6-DOF XFactory，保存原文命名差异。" } ], "verification": "verified", "directionsOriginal": [ "生成式装配", "物理约束规划" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Autonomously constructing physically realizable 3D structures remains a significant challenge due to combinatorial action spaces, interchangeable components, equifinal assembly sequences, and strict stability requirements during construction. State-of-the-art methods fine-tune large language models for text-based generative construction. However, these approaches do not allow for Multimodal (text, image, sketch) conditioning, overlook the practical role of scaffolding for stabilizing intermediate structures, and suffer from slow inference speeds. Therefore, we formulate construction as a probabilistic next-block generation task with multiple potential assembly actions and multiple potential task conditioning modalities. Concurrently, we explicitly consider the utility of scaffolding by introducing an auxiliary scaffold block token. We present Scaffold Multimodal Monte Carlo (ScaffoldM3C), a multimodal, lightweight, auto-regressive model for stable block-based construction, that proposes a set of next-step candidate blocks. Leveraging these candidates, we utilize Sequential Monte Carlo (SMC) to maintain a population of possible assembly sequences, allowing us to consider multiple, potentially different, assembly directions simultaneously. We train our multimodal architecture by extending the StableText2Brick dataset to contain image conditioning prompts and scaffold-stabilized build sequences. ScaffoldM3C is 4x smaller than competing baselines, yielding a 5x to 20x speedup during inference, while achieving comparable construction quality to state-of-the-art methods and higher overall stability. We demonstrate the effectiveness of our approach through simulations and real-world robot assembly demonstrations." }, { "id": "arxiv-2610.00524", "title": "Same Scene, Different Task: Skill Alignment for Compositional Generalization in VLAs", "titleZh": "同一场景，不同任务：通过技能对齐提升视觉语言动作模型的组合泛化", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00524", "paperUrl": "https://arxiv.org/abs/2610.00524", "projectUrl": null, "codeUrl": null, "category": "VLA与操作泛化", "tags": [ "反事实训练", "技能组合", "语言对齐", "真机实验" ], "directions": [ "视觉语言动作", "模仿学习" ], "robotFilters": [ "AgileX PiPER" ], "tier": "recent", "summary": "CRAFT在同图不同指令的反事实配对中对齐技能表示，Piper真机未示范组合成功43/60，常规微调为9/60。", "abstractZh": "论文针对微调时模型把视觉场景当作指令捷径的问题，保留同一观察但替换技能组合指令。新组合没有动作标签，因此将技能表征与视觉状态表征分开，从另一段已示范的同技能执行中转移监督，而不直接复制动作。三种VLA与两个仿真基准以及Piper真机结果表明，这能重组已见技能，但不解决全新技能或可变操作顺序。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "both", "experimentNote": "仿真Pick-Place含16组合仅示范4种，Pick-Place-Press含8组合仅示范2种，分别测π0、π0.5、GR00T N1.7。每组合50初始状态、三个评测种子；Piper真机每已见组合20演示，每组合测试5次。", "robots": [ "Piper 6-DoF（真机）", "Pick-Place/Press仿真机械臂" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "需复现VLM轨迹分段标签、反事实只改当前/未来操作的规则、三种损失及停止梯度位置，并严格分离示范与未示范组合。代码和新基准发布承诺已见于论文，本次未独立核验下载状态。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "增加技能与状态查询词元，前者读取图文，后者只读图像。以动作流匹配误差对比同技能/异技能表示，再将反事实技能表示与参考演示状态配对，监督参考动作速度场；分开路由梯度，避免直接拿异场景动作作标签。", "whyUseful": "需复现VLM轨迹分段标签、反事实只改当前/未来操作的规则、三种损失及停止梯度位置，并严格分离示范与未示范组合。代码和新基准发布承诺已见于论文，本次未独立核验下载状态。", "limitations": "要求所有组成技能已有演示，操作顺序固定；不能泛化成任意新技能规划。仿真误差条来自同一训练检查点的评测种子，未覆盖训练方差；真机每组合仅五次且场景较简单。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00524v1", "note": "PDF第7页表1：三个种子为同一训练检查点的环境/动作采样评测，而非三次独立训练。" }, { "url": "https://arxiv.org/pdf/2610.00524v1", "note": "PDF第9页表3与图6（已渲染核看）：Piper 6-DoF实际执行，未见组合43/60对9/60。" }, { "url": "https://arxiv.org/pdf/2610.00524v1", "note": "PDF第9页§6：固定操作序列与全部组成技能被演示为方法边界。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:57:30.176965+00:00", "original": { "id": "arxiv-2610.00524", "originalSourceUrl": "https://arxiv.org/pdf/2610.00524v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00524v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00524", "sha256": "75c9d2e3c4114da3eabfcf4745ed96375d8586a16425bdc0f777ea81385be618", "bytes": 13942404, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:44:22.189792+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00524", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00524", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00524", "pages": 26, "sourceIdentityStatus": "title_match", "sourceTitle": "Same Scene, Different Task: Skill Alignment for Compositional Generalization in VLAs", "sourceVersion": "2610.00524v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.741011+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00524" }, "originalAnalysis": { "id": "arxiv-2610.00524", "sectionsRead": [ "第3–4节问题与CRAFT，PDF第3–6页", "第5节实验、表1–3与消融，PDF第6–9页", "第6节讨论局限，PDF第9页" ], "analyzedAt": "2026-10-04T13:57:30.176965+00:00", "analysisStatus": "full_text_sections", "methodsZh": "增加技能与状态查询词元，前者读取图文，后者只读图像。以动作流匹配误差对比同技能/异技能表示，再将反事实技能表示与参考演示状态配对，监督参考动作速度场；分开路由梯度，避免直接拿异场景动作作标签。", "experimentsZh": "仿真Pick-Place含16组合仅示范4种，Pick-Place-Press含8组合仅示范2种，分别测π0、π0.5、GR00T N1.7。每组合50初始状态、三个评测种子；Piper真机每已见组合20演示，每组合测试5次。", "resultsZh": "π0.5在两基准未见组合成功率84.2%和62.6%，普通全微调为8.4%和0.7%。真机未见组合43/60，对照9/60；已见组合双方均19/20。仅技能对比不足，改变当前技能与保留当前技能的两类反事实监督共同贡献最大。", "limitationsZh": "要求所有组成技能已有演示，操作顺序固定；不能泛化成任意新技能规划。仿真误差条来自同一训练检查点的评测种子，未覆盖训练方差；真机每组合仅五次且场景较简单。", "reproductionZh": "需复现VLM轨迹分段标签、反事实只改当前/未来操作的规则、三种损失及停止梯度位置，并严格分离示范与未示范组合。代码和新基准发布承诺已见于论文，本次未独立核验下载状态。", "sourceUrl": "https://arxiv.org/pdf/2610.00524v1", "sourceVersion": "2610.00524v1", "originalSha256": "75c9d2e3c4114da3eabfcf4745ed96375d8586a16425bdc0f777ea81385be618", "experimentType": "both", "robots": [ "Piper 6-DoF（真机）", "Pick-Place/Press仿真机械臂" ], "evidenceNotes": [ { "section": "PDF第7页表1", "note": "三个种子为同一训练检查点的环境/动作采样评测，而非三次独立训练。" }, { "section": "PDF第9页表3与图6（已渲染核看）", "note": "Piper 6-DoF实际执行，未见组合43/60对9/60。" }, { "section": "PDF第9页§6", "note": "固定操作序列与全部组成技能被演示为方法边界。" } ], "verification": "verified", "directionsOriginal": [ "组合泛化", "视觉语言动作模型" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Vision-language-action (VLA) models often struggle to generalize to skill combinations absent from their fine-tuning demonstrations, even when every constituent skill has been demonstrated. We focus on a vision shortcut as one failure mode: during fine-tuning, visual observations can serve as a proxy for the instruction, so a policy may execute a demonstrated combination associated with similar observations rather than the instructed combination. This motivates training with counterfactual pairs formed by holding a demonstration observation fixed while changing the instruction to specify an undemonstrated combination. These pairs, however, lack corresponding demonstrated action targets. Crucially, the currently required skill has already been demonstrated, but actions from those executions cannot serve as direct targets because the same skill can require different actions across observations. We propose CRAFT, which transfers supervision from demonstrated executions of the required skill to counterfactual pairs using skill representations that can be reused across executions of the same skill. Across three VLA models and two simulation benchmarks, CRAFT improves success on undemonstrated combinations while maintaining high success on demonstrated ones; it also improves compositional generalization on a real robot. Project website: https://taegeunyang.github.io/craft/" }, { "id": "arxiv-2610.00542", "title": "Does Continual Imitation Learning Remain Grounded? A Language-Perturbed Benchmark for Robotic Task Retention", "titleZh": "持续模仿学习是否仍然具有语言依据？面向机器人任务保持的语言扰动基准", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00542", "paperUrl": "https://arxiv.org/abs/2610.00542", "projectUrl": null, "codeUrl": null, "category": "评测与数据集", "tags": [ "持续学习", "语言对齐", "指令扰动", "仿真" ], "directions": [ "模仿学习", "数据集与基准" ], "robotFilters": [], "tier": "recent", "summary": "通过改写、目标切换与不相容指令诊断持续学习，显示保留技能并不保证语言仍能正确控制行为。", "abstractZh": "该研究把任务保持能力与语言依赖程度分开评估：同义改写应保持行为，单语义槽改动应切换到新的有效目标，不可执行指令则用行为统计而非成功率评价。协议为LIBERO四套任务生成指令，但政策实验主要在LIBERO-Goal上比较八种持续模仿学习方法；并未测试大规模VLA。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "八种方法共用CLIP编码器、六层Transformer和混合高斯动作头，LIBERO-Goal十任务依次训练，每任务10轮、三个种子。每指令20个初始状态；改写覆盖各阶段，最小对比和冲突只测最终检查点。", "robots": [ "LIBERO-Goal机械臂（MuJoCo仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现需保存已验证指令清单、各方法任务ID/路由使用、配对初始状态及成功谓词。不可把无法执行的请求当作普通失败任务，也应同时报告绝对能力与改写差距，避免低能力模型因差距小被误判鲁棒。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "从任务定义抽取动作、对象、关系、目标、属性五类语义槽，生成同义改写、单槽有效对比与场景冲突请求，随后校验实体和目标谓词。每学习一项任务后冻结评估，结合成功AUC、改写差距、目标切换与旧目标持续率。", "whyUseful": "复现需保存已验证指令清单、各方法任务ID/路由使用、配对初始状态及成功谓词。不可把无法执行的请求当作普通失败任务，也应同时报告绝对能力与改写差距，避免低能力模型因差距小被误判鲁棒。", "limitations": "只测共同行为克隆架构，不能直接外推大VLA；最终阶段诊断无法分离原有缺陷与持续学习影响。正文对部分AUC及差距的叙述与表I不一致，故保留表格数值并标注待澄清。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00542v1", "note": "PDF第5页表I（已渲染核看）：DMPEL表中78.33/70.28/8.05，正文却写78.51/71.24/7.27，原文存在内部数字冲突。" }, { "url": "https://arxiv.org/pdf/2610.00542v1", "note": "PDF第5页表II：GSA要求达到新目标且不满足原目标，区别于一般target success。" }, { "url": "https://arxiv.org/pdf/2610.00542v1", "note": "PDF第8页局限：明确未评测large-scale VLAs；协议四套覆盖不等于四套政策实验。" } ], "verification": "partial", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:57:30.176472+00:00", "original": { "id": "arxiv-2610.00542", "originalSourceUrl": "https://arxiv.org/pdf/2610.00542v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00542v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00542", "sha256": "508b48f9d147d5f12c9dc7551df39add4e04b33d105c5938af30b5dedb7f6453", "bytes": 1746195, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:44:21.999371+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00542", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00542", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00542", "pages": 9, "sourceIdentityStatus": "title_match", "sourceTitle": "Does Continual Imitation Learning Remain Grounded? A Language-Perturbed Benchmark for Robotic Task Retention", "sourceVersion": "2610.00542v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.730720+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00542" }, "originalAnalysis": { "id": "arxiv-2610.00542", "sectionsRead": [ "第III节协议与指标，PDF第2–4页", "第IV–V节实验和表I–II，PDF第4–6页", "第VI节诊断分析，PDF第7页", "第VII节局限，PDF第7–8页" ], "analyzedAt": "2026-10-04T13:57:30.176472+00:00", "analysisStatus": "full_text_sections", "methodsZh": "从任务定义抽取动作、对象、关系、目标、属性五类语义槽，生成同义改写、单槽有效对比与场景冲突请求，随后校验实体和目标谓词。每学习一项任务后冻结评估，结合成功AUC、改写差距、目标切换与旧目标持续率。", "experimentsZh": "八种方法共用CLIP编码器、六层Transformer和混合高斯动作头，LIBERO-Goal十任务依次训练，每任务10轮、三个种子。每指令20个初始状态；改写覆盖各阶段，最小对比和冲突只测最终检查点。", "resultsZh": "按表I，DMPEL原始/改写AUC为78.33/70.28，最终成功率81.12%，却仍有8.05点改写差距。其有效新目标成功率30.91%，排除旧目标后切换率仅21.68%；不相容指令下多数方法仍启动动作。", "limitationsZh": "只测共同行为克隆架构，不能直接外推大VLA；最终阶段诊断无法分离原有缺陷与持续学习影响。正文对部分AUC及差距的叙述与表I不一致，故保留表格数值并标注待澄清。", "reproductionZh": "复现需保存已验证指令清单、各方法任务ID/路由使用、配对初始状态及成功谓词。不可把无法执行的请求当作普通失败任务，也应同时报告绝对能力与改写差距，避免低能力模型因差距小被误判鲁棒。", "sourceUrl": "https://arxiv.org/pdf/2610.00542v1", "sourceVersion": "2610.00542v1", "originalSha256": "508b48f9d147d5f12c9dc7551df39add4e04b33d105c5938af30b5dedb7f6453", "experimentType": "sim", "robots": [ "LIBERO-Goal机械臂（MuJoCo仿真）" ], "evidenceNotes": [ { "section": "PDF第5页表I（已渲染核看）", "note": "DMPEL表中78.33/70.28/8.05，正文却写78.51/71.24/7.27，原文存在内部数字冲突。" }, { "section": "PDF第5页表II", "note": "GSA要求达到新目标且不满足原目标，区别于一般target success。" }, { "section": "PDF第8页局限", "note": "明确未评测large-scale VLAs；协议四套覆盖不等于四套政策实验。" } ], "verification": "partial", "directionsOriginal": [ "持续模仿学习", "语义鲁棒性评测" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Continual imitation learning evaluates whether a robot can learn new knowledge without forgetting previously learned skills. However, retaining task performance does not ensure the behavior remains grounded in language because policies may rely on scene cues, object associations, or memorized task structure. We introduce a benchmark protocol to study how language-guided behavior changes as robotic policies learn successive tasks. We construct meaning-preserving and meaning-changing instruction variants for the Goal, Spatial, Object, and Long suites of LIBERO. Policy experiments focus on LIBERO-Goal, evaluating Original and Paraphrase instructions after each continual-learning stage. We compare representative continual imitation learning methods under their original assumptions while separating task competence from language sensitivity. The proposed diagnostics complement standard learning and forgetting metrics by measuring semantic robustness, goal adaptation, and language sensitivity. Results show that strong continual-learning performance does not always translate to reliable language grounding, and our diagnostics help determine whether retained skills remain correctly guided by their instructions. Additional materials are available at https://sites.google.com/view/stillgrounded" }, { "id": "arxiv-2610.00575", "title": "Token-World: World Modeling in Vision-Language Model Token Space for Robot Manipulation", "titleZh": "Token-World：在视觉语言模型词元空间中进行机器人操作世界建模", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00575", "paperUrl": "https://arxiv.org/abs/2610.00575", "projectUrl": null, "codeUrl": null, "category": "世界模型", "tags": [ "紧凑表征", "VLM词元", "策略评测", "世界模型" ], "directions": [ "世界模型" ], "robotFilters": [ "Franka Research 3" ], "tier": "recent", "summary": "直接预测压缩VLM视觉词元，策略模拟成功率相关系数达0.794；真机数据主要用于预测评测，不能视为新策略真机闭环验证。", "abstractZh": "Token-World将策略输入特征与动力学状态空间分开：先把高维视觉词元压缩，再学习动作条件转移，最后还原成原策略可用的词元。动力学循环无需生成RGB图像，辅以时序缓存和循环记忆维持较长预测。论文验证特征保真、动作一致性与学习仿真器对参考政策表现的拟合，并强调压缩容量与动力学可学性间的权衡。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "使用RoboTwin五十任务共5000条成功/失败轨迹，以及Franka六任务600条成功轨迹；按9:1划分。比较IRASim、Ctrl-World、WorldGym；三类共享Qwen3-VL的StarVLA头在学习仿真器中闭环运行，以解码视频人工判成功。", "robots": [ "Franka Research 3（真实轨迹采集）", "RoboTwin机械臂（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "应统一输入特征、轨迹切分和动作块跨度，并单独核对50步RGB基线与16步本方法的推理预算。保留RGB解码仅作评测的界限，不能把不同表格的条件混为一组；项目声明将开源，实际仓库未核实。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "冻结Qwen3-VL，将108个2560维视觉词元逐词压成16维，不合并空间位置。独立语义VAE先学重建，后冻结编码解码器；因果时空Transformer结合GRU与滑动窗口，以流匹配及shortcut forcing预测紧凑状态和本体信息。", "whyUseful": "应统一输入特征、轨迹切分和动作块跨度，并单独核对50步RGB基线与16步本方法的推理预算。保留RGB解码仅作评测的界限，不能把不同表格的条件混为一组；项目声明将开源，实际仓库未核实。", "limitations": "策略都共享同一视觉语言底座，跨底座迁移未测；人工视频成功判断与少量任务限制结论。主要证据是学习仿真和轨迹预测，记录的真实机械臂演示不等于所提世界模型控制真机的成功率。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00575v1", "note": "PDF第4–5页§IV-A及图3：真实平台为Franka Research 3、Robotiq夹爪、双RealSense 435；文中清晰描述的是轨迹采集和世界模型评测。" }, { "url": "https://arxiv.org/pdf/2610.00575v1", "note": "PDF第7页图6（已渲染核看）：0.794是政策—任务组合的模拟/参考成功率相关性，非任务成功率。" }, { "url": "https://arxiv.org/pdf/2610.00575v1", "note": "PDF第7页表III：16维动作NMSE不是所有维度中最小，48维为0.0236；综合最优不应写成全指标最优。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:57:30.175974+00:00", "original": { "id": "arxiv-2610.00575", "originalSourceUrl": "https://arxiv.org/pdf/2610.00575v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00575v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00575", "sha256": "beff8246d721bffc0f97751dc91735f6f80e0f9ca2fcf2588187391292fd0e2b", "bytes": 5065985, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:44:16.490527+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00575", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00575", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00575", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "Token-World: World Modeling in Vision-Language Model Token Space for Robot Manipulation", "sourceVersion": "2610.00575v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.728560+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00575" }, "originalAnalysis": { "id": "arxiv-2610.00575", "sectionsRead": [ "第III节与图2，PDF第3–4页", "第IV节实验及表I–IV、图6，PDF第4–7页", "第V节结论与局限，PDF第7页" ], "analyzedAt": "2026-10-04T13:57:30.175974+00:00", "analysisStatus": "full_text_sections", "methodsZh": "冻结Qwen3-VL，将108个2560维视觉词元逐词压成16维，不合并空间位置。独立语义VAE先学重建，后冻结编码解码器；因果时空Transformer结合GRU与滑动窗口，以流匹配及shortcut forcing预测紧凑状态和本体信息。", "experimentsZh": "使用RoboTwin五十任务共5000条成功/失败轨迹，以及Franka六任务600条成功轨迹；按9:1划分。比较IRASim、Ctrl-World、WorldGym；三类共享Qwen3-VL的StarVLA头在学习仿真器中闭环运行，以解码视频人工判成功。", "resultsZh": "RoboTwin特征余弦相似度0.7714、动作相似度0.9439。与参考成功率的Pearson相关为0.794，Ctrl-World为0.583；每预测观测0.359秒。16维较更宽状态在大多动力学指标上更优，重建更好不保证预测更好。", "limitationsZh": "策略都共享同一视觉语言底座，跨底座迁移未测；人工视频成功判断与少量任务限制结论。主要证据是学习仿真和轨迹预测，记录的真实机械臂演示不等于所提世界模型控制真机的成功率。", "reproductionZh": "应统一输入特征、轨迹切分和动作块跨度，并单独核对50步RGB基线与16步本方法的推理预算。保留RGB解码仅作评测的界限，不能把不同表格的条件混为一组；项目声明将开源，实际仓库未核实。", "sourceUrl": "https://arxiv.org/pdf/2610.00575v1", "sourceVersion": "2610.00575v1", "originalSha256": "beff8246d721bffc0f97751dc91735f6f80e0f9ca2fcf2588187391292fd0e2b", "experimentType": "sim", "robots": [ "Franka Research 3（真实轨迹采集）", "RoboTwin机械臂（仿真）" ], "evidenceNotes": [ { "section": "PDF第4–5页§IV-A及图3", "note": "真实平台为Franka Research 3、Robotiq夹爪、双RealSense 435；文中清晰描述的是轨迹采集和世界模型评测。" }, { "section": "PDF第7页图6（已渲染核看）", "note": "0.794是政策—任务组合的模拟/参考成功率相关性，非任务成功率。" }, { "section": "PDF第7页表III", "note": "16维动作NMSE不是所有维度中最小，48维为0.0236；综合最优不应写成全指标最优。" } ], "verification": "verified", "directionsOriginal": [ "机器人世界模型", "模型驱动仿真" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "A common approach to world-model simulation for vision-language-action (VLA) systems is to predict future RGB observations and then re-encode them into policy inputs, introducing an indirect interface between simulation and downstream policy execution. We instead investigate whether world dynamics can be modeled in a compact, policy-oriented state derived from VLM visual tokens. A key challenge is that raw VLM visual tokens are high-dimensional, making efficient and accurate autoregressive dynamics modeling challenging. To address this, we introduce Token-World, an action-conditioned world model that compresses VLM features into a compact token state, learns future dynamics in this reduced space, and maps predicted states back to the original policy-facing representation for downstream use. Across manipulation benchmarks, Token-World improves open-loop feature fidelity and policy-action consistency over recent world-model simulators, with slower degradation over long rollout horizons. In closed-loop evaluation, its simulated policy performance correlates more strongly with reference policy performance than Ctrl-World ($r=0.794$ vs.\\ $0.583$), while requiring lower simulation latency. Ablations further show that compact-representation design and dimensionality substantially affect future-state prediction. Code will be available at https://chuyaofu.github.io/Token-World/." }, { "id": "arxiv-2610.00601", "title": "When Reasoning Helps Action: Monitoring and Steering Chain-of-Thought in Vision-Language-Action Policies", "titleZh": "推理何时有助于行动：监测与引导视觉语言动作策略中的思维链", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00601", "paperUrl": "https://arxiv.org/abs/2610.00601", "projectUrl": null, "codeUrl": null, "category": "VLA推理与可靠性", "tags": [ "运行时监控", "思维链", "自动驾驶", "仿真" ], "directions": [ "视觉语言动作" ], "robotFilters": [], "tier": "recent", "summary": "TRUST能纠正两类VLA的推理，但闭环收益依赖动作是否遵循推理：驾驶有改善，操作成功率无显著变化。", "abstractZh": "论文区分推理的可纠正性与行动可响应性。TRUST以离线标注的推理轨迹训练独立价值模型，监测部分前缀并在低可信时重排下一词候选，冻结原策略。在驾驶与机器人操作仿真中，语言正确性都提高，但只有驾驶体现稳定的行为收益。结论是可读推理不能自动充当可靠的安全控制接口。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "对Alpamayo 1.5驾驶策略使用888段AlpaSim闭环场景，另预先按基线误差定义300段困难集；对DeepThinkVLA在LIBERO-Plus评测1509回合。与等延迟Best-of-4/32比较，并做100回合推理替换。", "robots": [ "AlpaSim自动驾驶车辆（仿真）", "LIBERO-Plus机械臂（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现应保留按场景划分、矛盾轨迹筛选、模糊标签排除、阈值及β标定；困难集效果不可写成全量收益。需同时报告文字正确性、配对闭环指标和意图执行，不以单一语言指标替代安全结果。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "TRUST根据图像与部分推理前缀预测最终正确概率，用完整轨迹标签训练token级价值头。低于验证集阈值时，对前十个下一词候选做门控价值增广采样；不改VLA权重，也不读取其隐藏态。", "whyUseful": "复现应保留按场景划分、矛盾轨迹筛选、模糊标签排除、阈值及β标定；困难集效果不可写成全量收益。需同时报告文字正确性、配对闭环指标和意图执行，不以单一语言指标替代安全结果。", "limitations": "正确性依赖VLM裁判，独立复判只覆盖部分场景；引导后前缀可能偏离价值训练分布。驾驶推理耗时约增81%；操作语义修改与动作变化不稳定，没有真机安全保障证据。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00601v1", "note": "PDF第5页表II（已渲染核看）：30.4%碰撞下降仅适用300段困难集；全888段下降5.6%，未标显著。" }, { "url": "https://arxiv.org/pdf/2610.00601v1", "note": "PDF第6页表III及闭环结果：语言纠正与任务成功改善分离，1509回合净成功减少9次。" }, { "url": "https://arxiv.org/pdf/2610.00601v1", "note": "PDF第7页表IV：ground-truth、语义翻转、乱码和去CoT的配对成功差异区间均含零。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:53:31.183167+00:00", "original": { "id": "arxiv-2610.00601", "originalSourceUrl": "https://arxiv.org/pdf/2610.00601v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00601v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00601", "sha256": "19a733df1a80ccf86a5b9beeacdd41c33ea32e6b4686f1892c4abb7182ae04df", "bytes": 2621149, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:44:11.840669+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00601", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00601", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00601", "pages": 10, "sourceIdentityStatus": "title_match", "sourceTitle": "When Reasoning Helps Action: Monitoring and Steering Chain-of-Thought in Vision-Language-Action Policies", "sourceVersion": "2610.00601v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.723969+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00601" }, "originalAnalysis": { "id": "arxiv-2610.00601", "sectionsRead": [ "第III节方法，PDF第2–3页", "第IV节实验设置，PDF第3–4页", "第V节结果与干预分析、表I–IV，PDF第4–7页", "第VI节讨论及附录A开头，PDF第8页" ], "analyzedAt": "2026-10-04T13:53:31.183167+00:00", "analysisStatus": "full_text_sections", "methodsZh": "TRUST根据图像与部分推理前缀预测最终正确概率，用完整轨迹标签训练token级价值头。低于验证集阈值时，对前十个下一词候选做门控价值增广采样；不改VLA权重，也不读取其隐藏态。", "experimentsZh": "对Alpamayo 1.5驾驶策略使用888段AlpaSim闭环场景，另预先按基线误差定义300段困难集；对DeepThinkVLA在LIBERO-Plus评测1509回合。与等延迟Best-of-4/32比较，并做100回合推理替换。", "resultsZh": "驾驶推理正确率75.9%升至90.0%；困难集碰撞率15.33%降至10.67%，相对下降30.4%，全量仅16.22%降至15.32%。操作抓取状态判断69.3%升至90.2%，成功率却78.86%变为78.26%，无显著提升。", "limitationsZh": "正确性依赖VLM裁判，独立复判只覆盖部分场景；引导后前缀可能偏离价值训练分布。驾驶推理耗时约增81%；操作语义修改与动作变化不稳定，没有真机安全保障证据。", "reproductionZh": "复现应保留按场景划分、矛盾轨迹筛选、模糊标签排除、阈值及β标定；困难集效果不可写成全量收益。需同时报告文字正确性、配对闭环指标和意图执行，不以单一语言指标替代安全结果。", "sourceUrl": "https://arxiv.org/pdf/2610.00601v1", "sourceVersion": "2610.00601v1", "originalSha256": "19a733df1a80ccf86a5b9beeacdd41c33ea32e6b4686f1892c4abb7182ae04df", "experimentType": "sim", "robots": [ "AlpaSim自动驾驶车辆（仿真）", "LIBERO-Plus机械臂（仿真）" ], "evidenceNotes": [ { "section": "PDF第5页表II（已渲染核看）", "note": "30.4%碰撞下降仅适用300段困难集；全888段下降5.6%，未标显著。" }, { "section": "PDF第6页表III及闭环结果", "note": "语言纠正与任务成功改善分离，1509回合净成功减少9次。" }, { "section": "PDF第7页表IV", "note": "ground-truth、语义翻转、乱码和去CoT的配对成功差异区间均含零。" } ], "verification": "verified", "directionsOriginal": [ "具身推理评估", "策略可靠性" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Reasoning-enabled VLA policies expose chain-of-thought (CoT) traces that appear to explain and guide their actions, creating a potential interface for runtime safety through reasoning monitoring and correction. In this work, we define and operationalize two evaluation axes for assessing when this interface can improve embodied behavior: correctability, which measures whether unreliable reasoning can be detected and improved during generation, and actionability, which measures whether reasoning corrections produce behaviorally meaningful changes in the intended direction. To enable correctability, we introduce Token-level Reward for Utility-Steered Chain-of-Thought (TRUST), an offline-trained value model that predicts eventual reasoning correctness from partial prefixes and uses these estimates to monitor and selectively steer reasoning generation in frozen VLA policies. On the Alpamayo 1.5 driving VLA, TRUST monitors correctness with 88.9% accuracy and improves reasoning correctness from 75.9% to 90.0%. On a baseline-defined challenging subset in AlpaSim, TRUST reduces collision rate by 30.4% and maximum trajectory error by 11.5% relative to the unsteered policy, outperforming a compute-matched Best-of-4 baseline. On the DeepThinkVLA manipulation VLA, TRUST improves the correctness of grasp-state claims from 69.3% to 90.2% and action-choice claims from 68.8% to 85.9%, yet closed-loop task performance on LIBERO-Plus remains largely unchanged. Empirical analysis reveals intent-consistent behavioral effects in Alpamayo 1.5 but limited effects in DeepThinkVLA, helping interpret these different task-level outcomes. Together, our results show that gains in reasoning correctness do not automatically imply gains in embodied performance, motivating evaluation of correctability and actionability when using CoT as a runtime safety interface." }, { "id": "arxiv-2610.00604", "title": "MIKASA-Robo-VLA: Benchmarking Memory in VLA Models for Long-Horizon Manipulation", "titleZh": "MIKASA-Robo-VLA：面向长时程操作的视觉语言动作模型记忆能力基准", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00604", "paperUrl": "https://arxiv.org/abs/2610.00604", "projectUrl": null, "codeUrl": null, "category": "评测与数据集", "tags": [ "记忆", "长时程操作", "VLA评测", "仿真" ], "directions": [ "视觉语言动作", "数据集与基准" ], "robotFilters": [ "Franka Panda" ], "tier": "recent", "summary": "提供90项语言条件记忆任务与22500条仿真轨迹；无显式记忆的π0.5在14项参考测试中成功率为21.1%。", "abstractZh": "该基准把必须记住的信息与单纯任务长度分开：80项任务隐藏后续动作依赖的线索，10项保留线索作为反应式对照，并按十类记忆需求组织。论文给出信息缺失间隔、统一观测动作接口、奖励和数据格式。有限基线显示困难，但不能把长任务失败单独归因于记忆不足。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "每任务收集250条成功oracle演示，总计22500条、约175.2小时仿真交互。π0.5仅用当前观测，在14项任务上全参数微调，一个训练运行和检查点，每任务20次不重叠种子测试，执行8步开环动作块。", "robots": [ "Franka Panda（ManiSkill 3仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "应明确种子、success_once口径及动作冻结包装；默认50回合与论文20回合协议不同。RLDS保留奖励和种子，LeRobot版本缺奖励且不能按episode索引跨格式连接；代码发布状态本次未独立验证。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "在ManiSkill 3中构造90项任务、26个家族和十类记忆需求，以线索出现、消失及动作阶段控制信息可用性。动作是10 Hz七维末端增量位姿与夹爪，观测为顶视、腕部图像和本体状态，另给稠密奖励与任务成功谓词。", "whyUseful": "应明确种子、success_once口径及动作冻结包装；默认50回合与论文20回合协议不同。RLDS保留奖励和种子，LeRobot版本缺奖励且不能按episode索引跨格式连接；代码发布状态本次未独立验证。", "limitations": "14项并非全基准代表样本，长时程与记忆类别、开环误差混杂。演示按成功筛选，不覆盖充分恢复；无真机迁移、无人工示范，语言模板有限，不能由此判定所有无记忆策略上限。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00604v1", "note": "PDF第6页§4.1：1.000 oracle成功率来自过滤成功回合，不是oracle未筛选可靠性。" }, { "url": "https://arxiv.org/pdf/2610.00604v1", "note": "PDF第8页表3（已渲染核看）：平均0.211，误差为跨任务标准误；仅14项任务。" }, { "url": "https://arxiv.org/pdf/2610.00604v1", "note": "PDF第9页§6：明确所有环境和演示均为仿真，迁移到物理机器人未测。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:53:31.182571+00:00", "original": { "id": "arxiv-2610.00604", "originalSourceUrl": "https://arxiv.org/pdf/2610.00604v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00604v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00604", "sha256": "a6eaf2e89521914bad5b86f1f05ea1caec20dd0690511a21730fe6a7a3f19b67", "bytes": 30283158, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:44:09.219187+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00604", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00604", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00604", "pages": 57, "sourceIdentityStatus": "title_match", "sourceTitle": "MIKASA-Robo-VLA: Benchmarking Memory in VLA Models for Long-Horizon Manipulation", "sourceVersion": "2610.00604v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.719406+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00604" }, "originalAnalysis": { "id": "arxiv-2610.00604", "sectionsRead": [ "第3–4节基准和数据，PDF第3–6页", "第5节分析及表3，PDF第6–9页", "第6–7节局限和结论，PDF第9页", "附录A的Panda动作与数据说明，PDF第14页（选段）" ], "analyzedAt": "2026-10-04T13:53:31.182571+00:00", "analysisStatus": "full_text_sections", "methodsZh": "在ManiSkill 3中构造90项任务、26个家族和十类记忆需求，以线索出现、消失及动作阶段控制信息可用性。动作是10 Hz七维末端增量位姿与夹爪，观测为顶视、腕部图像和本体状态，另给稠密奖励与任务成功谓词。", "experimentsZh": "每任务收集250条成功oracle演示，总计22500条、约175.2小时仿真交互。π0.5仅用当前观测，在14项任务上全参数微调，一个训练运行和检查点，每任务20次不重叠种子测试，执行8步开环动作块。", "resultsZh": "参考平均成功率0.211±0.044；长时程三项仅1/60成功。70项可测信息缺失间隔中28项超过16帧，但零间隔仍可能需要记忆；离散候选选择子集没有明确超越无提示猜测的证据。", "limitationsZh": "14项并非全基准代表样本，长时程与记忆类别、开环误差混杂。演示按成功筛选，不覆盖充分恢复；无真机迁移、无人工示范，语言模板有限，不能由此判定所有无记忆策略上限。", "reproductionZh": "应明确种子、success_once口径及动作冻结包装；默认50回合与论文20回合协议不同。RLDS保留奖励和种子，LeRobot版本缺奖励且不能按episode索引跨格式连接；代码发布状态本次未独立验证。", "sourceUrl": "https://arxiv.org/pdf/2610.00604v1", "sourceVersion": "2610.00604v1", "originalSha256": "a6eaf2e89521914bad5b86f1f05ea1caec20dd0690511a21730fe6a7a3f19b67", "experimentType": "sim", "robots": [ "Franka Panda（ManiSkill 3仿真）" ], "evidenceNotes": [ { "section": "PDF第6页§4.1", "note": "1.000 oracle成功率来自过滤成功回合，不是oracle未筛选可靠性。" }, { "section": "PDF第8页表3（已渲染核看）", "note": "平均0.211，误差为跨任务标准误；仅14项任务。" }, { "section": "PDF第9页§6", "note": "明确所有环境和演示均为仿真，迁移到物理机器人未测。" } ], "verification": "verified", "directionsOriginal": [ "记忆增强策略", "任务基准" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Vision-language-action policies often see only one or a few recent frames, which makes it difficult to evaluate how they use information that disappears during a task. We introduce MIKASA-Robo-VLA, a benchmark of 90 language-conditioned manipulation tasks. All but 10 hide the cue an action depends on. Those 10 are reactive controls. MIKASA-Robo, the suite it rebuilds, has 32 tasks and uses language only in a representative VLA subset. Here every task provides an instruction, while memory-dependent tasks hide a task-relevant cue and reactive controls keep it available. For 70 tasks, environment phase timings specify an information gap, and for 28 of them the gap exceeds the 16-frame window of the widest fixed-context VLA we survey. The gap counts only the interval the cue is provably absent, not the full duration a policy must retain it, so every memory-dependent task still requires memory by construction, including the ones whose measured gap is short. We release 22,500 oracle trajectories across 10 memory types in RLDS and LeRobotDataset v3. A reference $π_{0.5}$ baseline with current images and proprioception, but no observation history or explicit memory module, is fine-tuned on 14 tasks and achieves 0.211 $\\pm$ 0.044 mean task success. Its lower success on the evaluated Long-split tasks is confounded by open-loop chunking and the memory types represented in that subset. Project page: https://mikasarobo.github.io/" }, { "id": "arxiv-2610.00638", "title": "TacDyn-WAM: Learning Implicit Tactile Dynamics in a Heterogeneous Visuo-Tactile World Action Model", "titleZh": "TacDyn-WAM：在异构视触觉世界动作模型中学习隐式触觉动力学", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00638", "paperUrl": "https://arxiv.org/abs/2610.00638", "projectUrl": null, "codeUrl": null, "category": "触觉与接触操作", "tags": [ "视触觉", "世界动作模型", "隐式动力学", "真机实验" ], "directions": [ "世界模型" ], "robotFilters": [ "Franka Research 3" ], "tier": "recent", "summary": "通过动态触觉表征预测未来接触变化，仿真成功率81.5%，Franka五项真机任务经预训练达到85.0%。", "abstractZh": "本文把未来触觉预测从像素重建改为动态表征预测。TacRep以时空遮掩预测和局部关系蒸馏学习接触变化；独立触觉专家同时预测多个时间尺度的未来状态与变化量，并与视觉、动作专家联合注意。另一路只读触觉记忆提供当前接触信息。结果支持动态表征与当前触觉互补，但不是跨传感器通用能力的证明。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "both", "experimentNote": "UniVTAC八个仿真任务各100次回合；Franka Research 3搭载双Xense触觉传感器，五个真机任务各60条演示、20次测试。预训练使用6000条OmniViTac轨迹，并在触觉阶段加入300条目标任务演示。", "robots": [ "Franka Research 3（真机）", "UniVTAC平行夹爪（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现需保存背景差分和触觉归一化、四阶段冻结与注意掩码设置，并区分目标演示参与表征预训练的阶段。附录给出8张H100预训练、4张H100微调及逐阶段超参；本次未独立核实代码发布。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "以InternVLA-A1为底座，TacRep用四帧触觉视频作掩码时空预测，并以DINOv2局部关系蒸馏保留空间结构。0.4B触觉专家一次预测5、15、25帧后的表征及增量；AnyTouch2只读记忆补充当前接触，四阶段逐步开放联合注意。", "whyUseful": "复现需保存背景差分和触觉归一化、四阶段冻结与注意掩码设置，并区分目标演示参与表征预训练的阶段。附录给出8张H100预训练、4张H100微调及逐阶段超参；本次未独立核实代码发布。", "limitations": "每个配置只覆盖一种触觉传感器类型，跨传感器及非图像触觉泛化未测。HDMI插入仅12%，少量接触变化的数据限制优势；真机每任务20次且部分基线引用公开结果。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00638v1", "note": "PDF第7页表1：UniVTAC为仿真，81.5%并非超过两个大规模预训练N0模型。" }, { "url": "https://arxiv.org/pdf/2610.00638v1", "note": "PDF第9页表3–4（已渲染核看）：真机71.0%/85.0%，20次/任务；484 ms是50动作块而非单控制步。" }, { "url": "https://arxiv.org/pdf/2610.00638v1", "note": "PDF第8页§3.4：预训练前两阶段加入300条目标真机演示并过采样，不是纯外部数据预训练。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:53:31.182203+00:00", "original": { "id": "arxiv-2610.00638", "originalSourceUrl": "https://arxiv.org/pdf/2610.00638v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00638v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00638", "sha256": "c1ef04fd0a97fc2e91cd8b15a00b3ab452d9e2a036ce645b56bb35e93c4ed112", "bytes": 4839863, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:44:08.516208+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00638", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00638", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00638", "pages": 21, "sourceIdentityStatus": "title_match", "sourceTitle": "TacDyn-WAM: Learning Implicit Tactile Dynamics in a Heterogeneous Visuo-Tactile World Action Model", "sourceVersion": "2610.00638v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.690647+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00638" }, "originalAnalysis": { "id": "arxiv-2610.00638", "sectionsRead": [ "第2节方法，PDF第2–6页", "第3节实验及表1–4，PDF第7–9页", "第4节局限，PDF第9页", "附录D.2与表5–6，PDF第18页" ], "analyzedAt": "2026-10-04T13:53:31.182203+00:00", "analysisStatus": "full_text_sections", "methodsZh": "以InternVLA-A1为底座，TacRep用四帧触觉视频作掩码时空预测，并以DINOv2局部关系蒸馏保留空间结构。0.4B触觉专家一次预测5、15、25帧后的表征及增量；AnyTouch2只读记忆补充当前接触，四阶段逐步开放联合注意。", "experimentsZh": "UniVTAC八个仿真任务各100次回合；Franka Research 3搭载双Xense触觉传感器，五个真机任务各60条演示、20次测试。预训练使用6000条OmniViTac轨迹，并在触觉阶段加入300条目标任务演示。", "resultsZh": "仿真平均成功率81.5%，较底座提高25.7个百分点；去掉触觉世界模型或记忆降至67.9%和71.3%。真机无额外预训练为71.0%，预训练后85.0%。单A100上50步动作块延迟484毫秒。", "limitationsZh": "每个配置只覆盖一种触觉传感器类型，跨传感器及非图像触觉泛化未测。HDMI插入仅12%，少量接触变化的数据限制优势；真机每任务20次且部分基线引用公开结果。", "reproductionZh": "复现需保存背景差分和触觉归一化、四阶段冻结与注意掩码设置，并区分目标演示参与表征预训练的阶段。附录给出8张H100预训练、4张H100微调及逐阶段超参；本次未独立核实代码发布。", "sourceUrl": "https://arxiv.org/pdf/2610.00638v1", "sourceVersion": "2610.00638v1", "originalSha256": "c1ef04fd0a97fc2e91cd8b15a00b3ab452d9e2a036ce645b56bb35e93c4ed112", "experimentType": "both", "robots": [ "Franka Research 3（真机）", "UniVTAC平行夹爪（仿真）" ], "evidenceNotes": [ { "section": "PDF第7页表1", "note": "UniVTAC为仿真，81.5%并非超过两个大规模预训练N0模型。" }, { "section": "PDF第9页表3–4（已渲染核看）", "note": "真机71.0%/85.0%，20次/任务；484 ms是50动作块而非单控制步。" }, { "section": "PDF第8页§3.4", "note": "预训练前两阶段加入300条目标真机演示并过采样，不是纯外部数据预训练。" } ], "verification": "verified", "directionsOriginal": [ "多模态具身模型", "接触丰富操作" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "World action models improve robotic manipulation by conditioning actions on predicted futures, yet existing tactile variants largely inherit video-generation pipelines that reconstruct future tactile observations through iterative denoising. Such prediction can become unreliable under deployment drift: small changes in contact position or force may substantially alter tactile pixels even when the underlying contact evolution remains predictable. We introduce TacDyn-WAM, a heterogeneous visuo-tactile world action model that predicts implicit tactile dynamics rather than reconstructing future tactile observations. It learns TacRep, a dynamics-aware tactile target space trained through masked spatio-temporal prediction on tactile clips and regularized by relational structure distillation. A visual expert and an Implicit Tactile Dynamics Expert predict future visual and tactile representations in separate target spaces while interacting through joint attention; the tactile expert predicts future representations and their changes at multiple horizons in a single forward pass, and a read-only tactile memory supplies the current tactile state. On UniVTAC, TacDyn-WAM achieves an average success rate of 81.5% using only the provided demonstrations, reaching state-of-the-art-level performance and remaining competitive with models pretrained on large-scale visuo-tactile trajectories. Ablations confirm the benefits of both tactile pathways and TacRep over pixel-reconstruction and static alternatives. On five real-robot tasks, TacDyn-WAM reaches 71.0% average success, and modest-scale pretraining raises it to 85.0%, further validating our method." }, { "id": "arxiv-2610.00718", "title": "Toward Humanoid Robots in Construction: A Teleoperation Feasibility Study", "titleZh": "迈向建筑施工中的人形机器人：一项遥操作可行性研究", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00718", "paperUrl": "https://arxiv.org/abs/2610.00718", "projectUrl": null, "codeUrl": null, "category": "人形机器人应用", "tags": [ "建筑施工", "XR遥操作", "全身操作", "现场验证" ], "directions": [ "运动控制" ], "robotFilters": [ "Unitree G1", "Inspire五指灵巧手" ], "tier": "recent", "summary": "XR手部控制加脚踏行走，使坐姿操作员在工地完成搬工具和滚刷涂覆。", "abstractZh": "系统集成Quest头手追踪、主动双目视角与脚踏移动，控制搭载灵巧手的Unitree G1。两名操作员在实际工地完成两种可行性任务，成功率较高，但速度远慢于人工并出现抓握不稳和电机过热。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "real", "experimentNote": "在活跃建筑工地使用Unitree G1和带触觉的Inspire五指手。两名操作员分别具有/不具有VR经验，各任务各10次；搬运手工具约4米到箱中，滚刷在纸面覆盖至少90%。另记录同一操作员人工完成时间。", "robots": [ "Unitree G1（实机）", "Inspire五指灵巧手（实机）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现应明确脚踏映射、颈部控制和上下肢命令协调，保留急停与热管理，并按原成功定义计时。应扩展操作者和现场条件、报告重置与准备时间；记录可用于未来模仿学习不代表已训练或验证自主施工策略。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "Quest 3将头部俯仰/偏航映射到两自由度ZED Mini相机颈部，通过WebRTC回传双目视频；Unitree XR框架重定向上肢与手指。GLYDR脚踏控制行走、转向和蹲起，GR00T Whole Body Control产生下肢命令，使双手操作与移动并行。", "whyUseful": "复现应明确脚踏映射、颈部控制和上下肢命令协调，保留急停与热管理，并按原成功定义计时。应扩展操作者和现场条件、报告重置与准备时间；记录可用于未来模仿学习不代表已训练或验证自主施工策略。", "limitations": "仅两操作者、两任务的初步可行性，不能推为施工生产率或自主能力。任务受载荷和可抓取范围限制，涂刷失败主要来自滚刷抓握不稳；长时间遥操作出现电机过热，尚未量化工效学收益与持续运行安全。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00718v1", "note": "§III，PDF页2：Quest 3、GLYDR、G1、Inspire与GR00T全身控制，数据为LeRobot格式。" }, { "url": "https://arxiv.org/pdf/2610.00718v1", "note": "表II，PDF页4：每任务2人×10次；搬运72秒/4秒、涂刷149秒/47秒。" }, { "url": "https://arxiv.org/pdf/2610.00718v1", "note": "§IV–V，PDF页3–4：实地遥操作而非自主；抓握及电机过热是实际观测到的限制。" }, { "url": "https://arxiv.org/pdf/2610.00718v1", "note": "PDF页4（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823254+00:00", "original": { "id": "arxiv-2610.00718", "originalSourceUrl": "https://arxiv.org/pdf/2610.00718v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00718v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00718", "sha256": "aaa282158ac3d1baeefcf3616f562b5c9865fa61876280a0e2ae96fbd7835ce1", "bytes": 22118302, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:58.036533+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00718", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00718", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00718", "pages": 4, "sourceIdentityStatus": "title_match", "sourceTitle": "Toward Humanoid Robots in Construction: A Teleoperation Feasibility Study", "sourceVersion": "2610.00718v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.685419+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00718" }, "originalAnalysis": { "id": "arxiv-2610.00718", "sectionsRead": [ "§III，PDF页2", "§IV及表I，PDF页2–3", "表II及§V，PDF页4" ], "analyzedAt": "2026-10-04T13:58:28.823254+00:00", "analysisStatus": "full_text_sections", "methodsZh": "Quest 3将头部俯仰/偏航映射到两自由度ZED Mini相机颈部，通过WebRTC回传双目视频；Unitree XR框架重定向上肢与手指。GLYDR脚踏控制行走、转向和蹲起，GR00T Whole Body Control产生下肢命令，使双手操作与移动并行。", "experimentsZh": "在活跃建筑工地使用Unitree G1和带触觉的Inspire五指手。两名操作员分别具有/不具有VR经验，各任务各10次；搬运手工具约4米到箱中，滚刷在纸面覆盖至少90%。另记录同一操作员人工完成时间。", "resultsZh": "搬工具20/20成功，均时72秒，人工4秒，约慢18倍；滚刷16/20成功，149秒对47秒，约慢3倍。VR熟练者两任务更快。系统同步记录相机、触觉、电机状态及命令为LeRobot格式。", "limitationsZh": "仅两操作者、两任务的初步可行性，不能推为施工生产率或自主能力。任务受载荷和可抓取范围限制，涂刷失败主要来自滚刷抓握不稳；长时间遥操作出现电机过热，尚未量化工效学收益与持续运行安全。", "reproductionZh": "复现应明确脚踏映射、颈部控制和上下肢命令协调，保留急停与热管理，并按原成功定义计时。应扩展操作者和现场条件、报告重置与准备时间；记录可用于未来模仿学习不代表已训练或验证自主施工策略。", "sourceUrl": "https://arxiv.org/pdf/2610.00718v1", "sourceVersion": "2610.00718v1", "originalSha256": "aaa282158ac3d1baeefcf3616f562b5c9865fa61876280a0e2ae96fbd7835ce1", "experimentType": "real", "robots": [ "Unitree G1（实机）", "Inspire五指灵巧手（实机）" ], "evidenceNotes": [ { "section": "§III，PDF页2", "note": "Quest 3、GLYDR、G1、Inspire与GR00T全身控制，数据为LeRobot格式。" }, { "section": "表II，PDF页4", "note": "每任务2人×10次；搬运72秒/4秒、涂刷149秒/47秒。" }, { "section": "§IV–V，PDF页3–4", "note": "实地遥操作而非自主；抓握及电机过热是实际观测到的限制。" }, { "section": "PDF页4（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "施工人形机器人", "坐姿遥操作采集" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "We present a teleoperation system that enables a single operator to perform construction tasks on a Unitree G1 humanoid, combining extended reality (XR) based upper body control with pedal-based locomotion to enable simultaneous manipulation and locomotion. Motivated by persistent labor shortages, hazardous working conditions, and challenges in humanoid autonomy, we investigate teleoperation as a practical near-term approach for reducing physical strain on workers while generating high quality demonstration data. We evaluate the system on two representative construction tasks drawn from O*NET occupational database, and report task success and completion time relative to a manual baseline. The system achieved 100% success on tool transport and 80% success on surface painting, with teleoperation requiring substantially more time compared to manual execution." }, { "id": "arxiv-2610.00727", "title": "CF-JEPA: Improving Robustness of JEPA World Models via Controllability Factorization", "titleZh": "CF-JEPA：通过可控性因子分解提高JEPA世界模型的鲁棒性", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00727", "paperUrl": "https://arxiv.org/abs/2610.00727", "projectUrl": null, "codeUrl": null, "category": "世界模型与表征学习", "tags": [ "JEPA", "可控性分解", "视觉干扰", "潜空间坍塌" ], "directions": [ "世界模型" ], "robotFilters": [], "tier": "recent", "summary": "把可控与不可控潜变量分开，减轻不相关视觉干扰造成的表征坍塌。", "abstractZh": "CF-JEPA在无奖励离线转移数据上学习两个潜空间，动作预测约束可控部分，梯度反转抑制不可控部分携带动作信息。规划只使用可控潜变量；研究同时揭示某些基准的动作接口会绕开当前视觉，鲁棒性评测需谨慎。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "四基础任务为Reacher、PushT、OGBench Cube和TwoRoom，加入不遮挡主体的边缘移动彩色矩形。再以25万帧训练关节位置控制的简化ManiSkill Reach，16干扰物下测不同精度阈值；模型通常3训练种子。", "robots": [ "Reacher（仿真）", "OGBench机械臂（仿真）", "ManiSkill机械臂（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现需相同离线轨迹、干扰生成和损失权重，区分训练时就加干扰与测试突发变化。CEM为300候选、30轮、5步规划。必须核验控制是否真正依赖当前视觉，OGBench末端IK接口可在遮挡下保持表现。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "ViT-Tiny输出192维潜变量，拆为128维可控与64维不可控。两支分别用动作条件/无动作动力学预测，加入可控支逆动力学、不可控支梯度反转动作预测和SIGReg防坍塌；CEM仅最小化可控潜状态到目标图的距离。", "whyUseful": "复现需相同离线轨迹、干扰生成和损失权重，区分训练时就加干扰与测试突发变化。CEM为300候选、30轮、5步规划。必须核验控制是否真正依赖当前视觉，OGBench末端IK接口可在遮挡下保持表现。", "limitations": "所有验证为仿真及单类人工干扰。假设不可控状态不影响可控动力学，不适合遮挡或人与机器人交互。干净任务并非普遍持平；名义基线部分直接引自原文，不能视为全条件统一重跑。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00727v1", "note": "表II–III，PDF页4–5：无干扰明显落后部分基线，有干扰条件改善；名义基线引用源论文。" }, { "url": "https://arxiv.org/pdf/2610.00727v1", "note": "§IV-C–D，PDF页6：OGBench当前视觉可被动作接口绕开，因此新增关节控制ManiSkill测试。" }, { "url": "https://arxiv.org/pdf/2610.00727v1", "note": "§V，PDF页7：单干扰族、仅仿真及可控/不可控独立假设。" }, { "url": "https://arxiv.org/pdf/2610.00727v1", "note": "PDF页6（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823252+00:00", "original": { "id": "arxiv-2610.00727", "originalSourceUrl": "https://arxiv.org/pdf/2610.00727v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00727v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00727", "sha256": "1ea7f2fb228bc0bf9412956b4af0dd51334cab1925308e528474e7b141467429", "bytes": 585069, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:57.458735+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00727", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00727", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00727", "pages": 9, "sourceIdentityStatus": "title_match", "sourceTitle": "CF-JEPA: Improving Robustness of JEPA World Models via Controllability Factorization", "sourceVersion": "2610.00727v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.667784+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00727" }, "originalAnalysis": { "id": "arxiv-2610.00727", "sectionsRead": [ "§III，PDF页2–5", "§IV-A–C及表II–IV，PDF页4–6", "§IV-D–E与表V–VI，PDF页6–7", "§V，PDF页7" ], "analyzedAt": "2026-10-04T13:58:28.823252+00:00", "analysisStatus": "full_text_sections", "methodsZh": "ViT-Tiny输出192维潜变量，拆为128维可控与64维不可控。两支分别用动作条件/无动作动力学预测，加入可控支逆动力学、不可控支梯度反转动作预测和SIGReg防坍塌；CEM仅最小化可控潜状态到目标图的距离。", "experimentsZh": "四基础任务为Reacher、PushT、OGBench Cube和TwoRoom，加入不遮挡主体的边缘移动彩色矩形。再以25万帧训练关节位置控制的简化ManiSkill Reach，16干扰物下测不同精度阈值；模型通常3训练种子。", "resultsZh": "两干扰物Reacher成功0.48，对LeWM0.07、SMWM0.10；64干扰物PushT为0.57对SMWM0.27。ManiSkill两厘米阈值0.35对0.22。但无干扰Reacher/PushT仅0.61/0.78，落后LeWM0.86/0.96。", "limitationsZh": "所有验证为仿真及单类人工干扰。假设不可控状态不影响可控动力学，不适合遮挡或人与机器人交互。干净任务并非普遍持平；名义基线部分直接引自原文，不能视为全条件统一重跑。", "reproductionZh": "复现需相同离线轨迹、干扰生成和损失权重，区分训练时就加干扰与测试突发变化。CEM为300候选、30轮、5步规划。必须核验控制是否真正依赖当前视觉，OGBench末端IK接口可在遮挡下保持表现。", "sourceUrl": "https://arxiv.org/pdf/2610.00727v1", "sourceVersion": "2610.00727v1", "originalSha256": "1ea7f2fb228bc0bf9412956b4af0dd51334cab1925308e528474e7b141467429", "experimentType": "sim", "robots": [ "Reacher（仿真）", "OGBench机械臂（仿真）", "ManiSkill机械臂（仿真）" ], "evidenceNotes": [ { "section": "表II–III，PDF页4–5", "note": "无干扰明显落后部分基线，有干扰条件改善；名义基线引用源论文。" }, { "section": "§IV-C–D，PDF页6", "note": "OGBench当前视觉可被动作接口绕开，因此新增关节控制ManiSkill测试。" }, { "section": "§V，PDF页7", "note": "单干扰族、仅仿真及可控/不可控独立假设。" }, { "section": "PDF页6（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "任务相关世界模型", "视觉控制鲁棒性" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Controlling an agent with vision requires being able to separate useful information from irrelevant background information. JEPA-style latent world models seem like a natural approach for this, as they do not perform pixel-level reconstruction; however, they are still sensitive to these distractor signals and experience latent collapse. In this work, we introduce Controllability Factorized JEPA (CF-JEPA), a JEPA-style world model which splits the latent space into controllable and uncontrollable subspaces. This factorization allows us to capture all the distractor information into the uncontrollable region, while we use the control-relevant latent information for our task. With this, we show comparable performance across 2D and 3D control tasks under nominal conditions and improved performance under distracted conditions, where CF-JEPA is the only model that does not experience latent collapse. We also validate our model under distracted conditions for a simulated robot task, highlighting the practical application of such a scheme." }, { "id": "arxiv-2610.00729", "title": "Reward as Observation: Learning Reward-Based Policies for Rapid Adaptation", "titleZh": "将奖励作为观测：学习基于奖励的策略以快速适应", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00729", "paperUrl": "https://arxiv.org/abs/2610.00729", "projectUrl": null, "codeUrl": null, "category": "强化学习与迁移", "tags": [ "奖励历史", "部分可观测", "零样本迁移", "DAgger" ], "directions": [ "强化学习", "模仿学习" ], "robotFilters": [ "Stretch" ], "tier": "recent", "summary": "只依赖奖励与动作历史的策略，可跨视觉外观迁移并指导目标域视觉策略。", "abstractZh": "该方法把观测替换为奖励—动作历史，用LSTM和专家监督训练PPO策略。只要相关动力学、动作及奖励结构保持一致，视觉变化不影响策略输入；作者还将其作为教师加速新图像域学习，物理硬件仍未测试。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "在Pointmass、DMC Cartpole及2D赛车训练，跨颜色变化、同动力学3D赛车渲染及Habitat-Sim两处HM3D真实建筑扫描测试。奖励由距离与航向密集信号构造；另用10万目标图像样本对比奖励和状态估计。", "robots": [ "Stretch移动机器人（Habitat-Sim仿真）", "DMC Cartpole（仿真）", "Pointmass及赛车代理（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现需保持奖励定义、动作尺度和动力学一致，明确颜色/渲染变化与真实动力学迁移不同。应计入源域专家及奖励估计数据成本，并测量探索步骤；不要把真实楼宇扫描或Stretch仿真渲染记作实体机器人试验。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "将原POMDP替换成奖励/动作历史输入，MLP-LSTM隐状态送入策略和价值网络。PPO之外加入模仿状态专家的动作损失以缓解探索困难；迁移后可用该策略给目标域访问状态提供DAgger标签，最终学生只读取目标图像。", "whyUseful": "复现需保持奖励定义、动作尺度和动力学一致，明确颜色/渲染变化与真实动力学迁移不同。应计入源域专家及奖励估计数据成本，并测量探索步骤；不要把真实楼宇扫描或Stretch仿真渲染记作实体机器人试验。", "limitations": "需要全程信息充分的密集奖励，稀疏奖励无法有效定位；源/目标共享动力学等是假设而非一般迁移保证。历史标量使高维探索更难，6维Pointmass仅约57%最优奖励；实机定位、奖励获取和动态差异未验证。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00729v1", "note": "§III–IV，PDF页2–4：相同状态、动作、转移、奖励而观测可变；依赖密集奖励。" }, { "url": "https://arxiv.org/pdf/2610.00729v1", "note": "§V-B图6与§VI，PDF页5–6：Stretch运行于Habitat-Sim真实楼宇重建，结论明确硬件未测试。" }, { "url": "https://arxiv.org/pdf/2610.00729v1", "note": "PDF页6（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823250+00:00", "original": { "id": "arxiv-2610.00729", "originalSourceUrl": "https://arxiv.org/pdf/2610.00729v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00729v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00729", "sha256": "90eee95f553908b718ce7662a436a4b991a3793a77802119df3501805e00e131", "bytes": 2916928, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:52.379876+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00729", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00729", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00729", "pages": 9, "sourceIdentityStatus": "title_match", "sourceTitle": "Reward as Observation: Learning Reward-Based Policies for Rapid Adaptation", "sourceVersion": "2610.00729v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.666183+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00729" }, "originalAnalysis": { "id": "arxiv-2610.00729", "sectionsRead": [ "§III–IV，PDF页2–4", "§V-A–C及表I–II，PDF页3–5", "§V-D–E及表III，PDF页5–6", "§VI，PDF页6" ], "analyzedAt": "2026-10-04T13:58:28.823250+00:00", "analysisStatus": "full_text_sections", "methodsZh": "将原POMDP替换成奖励/动作历史输入，MLP-LSTM隐状态送入策略和价值网络。PPO之外加入模仿状态专家的动作损失以缓解探索困难；迁移后可用该策略给目标域访问状态提供DAgger标签，最终学生只读取目标图像。", "experimentsZh": "在Pointmass、DMC Cartpole及2D赛车训练，跨颜色变化、同动力学3D赛车渲染及Habitat-Sim两处HM3D真实建筑扫描测试。奖励由距离与航向密集信号构造；另用10万目标图像样本对比奖励和状态估计。", "resultsZh": "奖励策略达到状态策略约95%、66%、70%表现。3D赛车学生约7万步接近最优；图像Cartpole中奖励估计路线回报735.83，状态估计路线189.41，后者近随机192.78。Habitat示例为仿真Stretch导航。", "limitationsZh": "需要全程信息充分的密集奖励，稀疏奖励无法有效定位；源/目标共享动力学等是假设而非一般迁移保证。历史标量使高维探索更难，6维Pointmass仅约57%最优奖励；实机定位、奖励获取和动态差异未验证。", "reproductionZh": "复现需保持奖励定义、动作尺度和动力学一致，明确颜色/渲染变化与真实动力学迁移不同。应计入源域专家及奖励估计数据成本，并测量探索步骤；不要把真实楼宇扫描或Stretch仿真渲染记作实体机器人试验。", "sourceUrl": "https://arxiv.org/pdf/2610.00729v1", "sourceVersion": "2610.00729v1", "originalSha256": "90eee95f553908b718ce7662a436a4b991a3793a77802119df3501805e00e131", "experimentType": "sim", "robots": [ "Stretch移动机器人（Habitat-Sim仿真）", "DMC Cartpole（仿真）", "Pointmass及赛车代理（仿真）" ], "evidenceNotes": [ { "section": "§III–IV，PDF页2–4", "note": "相同状态、动作、转移、奖励而观测可变；依赖密集奖励。" }, { "section": "§V-B图6与§VI，PDF页5–6", "note": "Stretch运行于Habitat-Sim真实楼宇重建，结论明确硬件未测试。" }, { "section": "PDF页6（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "跨观测空间迁移", "奖励驱动控制" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "This paper explores a reward-based policy to achieve zero-shot transfer between source and target environments with completely different observation spaces. While humans can demonstrate impressive adaptation capabilities, deep neural network policies often struggle to adapt to a new environment and require a considerable amount of samples for successful transfer. Instead, we propose a novel reward-based policy only conditioned on rewards and actions, enabling zero-shot adaptation to new environments with completely different observations. We discuss the challenges and feasibility of a reward-based policy and then propose a practical algorithm for training. We demonstrate that a reward policy can be trained within three different environments, Pointmass, Cartpole, and 2D Car Racing, and transferred to completely different observations, such as different color palettes or 3D rendering, or Stretch robot navigation in Habitat-Sim, in a zero-shot manner. We also demonstrate that a reward-based policy can further guide the training of an observation-based policy in the target environment." }, { "id": "arxiv-2610.00731", "title": "Measuring Asset and Scene Reconstruction Effects in Real-to-Sim Robot Evaluation", "titleZh": "衡量资产与场景重建对真实到仿真机器人评测的影响", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00731", "paperUrl": "https://arxiv.org/abs/2610.00731", "projectUrl": null, "codeUrl": null, "category": "仿真与真实评测", "tags": [ "Real-to-Sim", "环境重建", "物理参数", "评测一致性" ], "directions": [ "数据集与基准" ], "robotFilters": [ "YAM" ], "tier": "recent", "summary": "同一机器人单元重建两次，量化完整重建流程对仿真—实机分数差距的影响。", "abstractZh": "论文比较带度量尺度和定制物理参数的重建，与按公开配方制作的默认重建。固定机器人、策略及运行设置，对照第三方执行的实体试验；整体重建改进显著减少分数和任务进度差距，但不能分解单项资产属性的作用。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "both", "experimentNote": "I2RT YAM双臂工作站，π0.5与MolmoAct2各做5任务，形成10任务—策略单元。每重建每单元20次，共400仿真试验，与Robocurve付费委托执行并评分的200实机试验比较；仿真按同量表自动判分。", "robots": [ "I2RT YAM双臂工作站（实机与Isaac Sim）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现应锁定Isaac Sim 6.0.1、运行配置与原始每试验得分，按任务—策略单元配对而非逐回合配对。应重新抽样计算置信区间、保留失败单元和被门禁拒绝记录；不能把整个流程效果宣传为某一物理参数的因果贡献。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "两版本使用相同对象照片与场景视频。定制版结合ChArUco尺度、投影纹理、按物体选碰撞体、VLM估计质量/回弹和类别摩擦表；默认版用TRELLIS、人工目测尺度和PhysX默认参数。运行门禁核对相机、动作块、预算等9项配置。", "whyUseful": "复现应锁定Isaac Sim 6.0.1、运行配置与原始每试验得分，按任务—策略单元配对而非逐回合配对。应重新抽样计算置信区间、保留失败单元和被门禁拒绝记录；不能把整个流程效果宣传为某一物理参数的因果贡献。", "limitations": "几何、尺度、纹理、物理和场景同时改变，无法单因素归因。仅单场地两策略五任务；实机人工与仿真规则评分有差异，仿真初始位移更窄。默认配方由作者重建且部分输入弱于原方法，相关性提高未达差值显著。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00731v1", "note": "§3，PDF页3–4：400仿真与200第三方实机试验；作者雇主支付实体评测费用。" }, { "url": "https://arxiv.org/pdf/2610.00731v1", "note": "表4，PDF页11：分数/进度改善区间不跨零，相关系数/失败阶段改善区间跨零。" }, { "url": "https://arxiv.org/pdf/2610.00731v1", "note": "§6，PDF页14：多因素同时变化、评分差异和默认重建潜在偏向明确列为限制。" }, { "url": "https://arxiv.org/pdf/2610.00731v1", "note": "PDF页11（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823248+00:00", "original": { "id": "arxiv-2610.00731", "originalSourceUrl": "https://arxiv.org/pdf/2610.00731v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00731v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00731", "sha256": "392146b25295eb4489f0c3a6b922d8e576029442921eb014e489a7f6a5bb025e", "bytes": 3191185, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:47.771774+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00731", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00731", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00731", "pages": 20, "sourceIdentityStatus": "title_match", "sourceTitle": "Measuring Asset and Scene Reconstruction Effects in Real-to-Sim Robot Evaluation", "sourceVersion": "2610.00731v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.662798+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00731" }, "originalAnalysis": { "id": "arxiv-2610.00731", "sectionsRead": [ "§3及表1，PDF页3–5", "§4及表2，PDF页7–8", "§5及表4–6，PDF页11–13", "§6–7，PDF页14–15" ], "analyzedAt": "2026-10-04T13:58:28.823248+00:00", "analysisStatus": "full_text_sections", "methodsZh": "两版本使用相同对象照片与场景视频。定制版结合ChArUco尺度、投影纹理、按物体选碰撞体、VLM估计质量/回弹和类别摩擦表；默认版用TRELLIS、人工目测尺度和PhysX默认参数。运行门禁核对相机、动作块、预算等9项配置。", "experimentsZh": "I2RT YAM双臂工作站，π0.5与MolmoAct2各做5任务，形成10任务—策略单元。每重建每单元20次，共400仿真试验，与Robocurve付费委托执行并评分的200实机试验比较；仿真按同量表自动判分。", "resultsZh": "平均分数误差6.97对17.54百分点，改善10.56，95%区间5.03–17.11；进度差距改善8.32点。相关系数0.90对0.51，但差值置信区间跨零；10单元中7个明确更接近实机，3个视为并列。", "limitationsZh": "几何、尺度、纹理、物理和场景同时改变，无法单因素归因。仅单场地两策略五任务；实机人工与仿真规则评分有差异，仿真初始位移更窄。默认配方由作者重建且部分输入弱于原方法，相关性提高未达差值显著。", "reproductionZh": "复现应锁定Isaac Sim 6.0.1、运行配置与原始每试验得分，按任务—策略单元配对而非逐回合配对。应重新抽样计算置信区间、保留失败单元和被门禁拒绝记录；不能把整个流程效果宣传为某一物理参数的因果贡献。", "sourceUrl": "https://arxiv.org/pdf/2610.00731v1", "sourceVersion": "2610.00731v1", "originalSha256": "392146b25295eb4489f0c3a6b922d8e576029442921eb014e489a7f6a5bb025e", "experimentType": "both", "robots": [ "I2RT YAM双臂工作站（实机与Isaac Sim）" ], "evidenceNotes": [ { "section": "§3，PDF页3–4", "note": "400仿真与200第三方实机试验；作者雇主支付实体评测费用。" }, { "section": "表4，PDF页11", "note": "分数/进度改善区间不跨零，相关系数/失败阶段改善区间跨零。" }, { "section": "§6，PDF页14", "note": "多因素同时变化、评分差异和默认重建潜在偏向明确列为限制。" }, { "section": "PDF页11（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "仿真评测可信度", "数字孪生重建质量" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Simulated evaluation is increasingly used alongside real-world evaluation of robot policies because it is cheaper and easier to repeat; however, its value depends on how closely its outcomes track the real robot's. We test whether our reconstruction pipeline, combining metrically scaled object geometry, authored physical parameters and scene reconstruction, reduces disagreement between simulated and real robot scores relative to a default open-source recipe. We constructed two simulated versions of one bimanual robot cell: an authored reconstruction, using object geometry at estimated metric scale, projected textures, authored physics and our own scene splat; and a baseline, referred to as the default reconstruction, using the open-source recipe of a generative single-image mesh, engine-default physics and a Gaussian-splat scene. Both reconstructions use the same object photographs and scene video. Two policies ran five tasks each, giving ten task-policy pairs, which we call cells; each cell was run twenty times in each reconstruction with all other settings held fixed. Both reconstructions were scored against the same real trials, graded by a third-party evaluator. Pearson correlation between the ten simulated and real cell means is r = 0.90 for the authored reconstruction and 0.51 for the default. Mean score error is 6.97 percentage points for the authored reconstruction and 17.54 for the default, a reduction of 10.56 percentage points. These results show that improving the quality of the environment reconstruction through higher visual fidelity, authored physics and metric scale makes the simulation more faithful to the real world and narrows the sim-to-real gap. We release the harness, the per-trial scores, every reported run's configuration, and the assets and scenes of both reconstructions." }, { "id": "arxiv-2610.00781", "title": "DITTO-X: Forward and Reverse Teleoperation for Dexterous Manipulation and Human Intervention", "titleZh": "DITTO-X：用于灵巧操作与人类干预的正向和反向遥操作", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00781", "paperUrl": "https://arxiv.org/abs/2610.00781", "projectUrl": null, "codeUrl": null, "category": "遥操作与灵巧手", "tags": [ "力反馈", "触觉反馈", "反向遥操作", "DAgger" ], "directions": [ "模仿学习", "灵巧手" ], "robotFilters": [ "Franka Panda", "Sharpa Wave", "Wuji 2", "Inspire RH56DFX" ], "tier": "recent", "summary": "用主动外骨骼将人手预先对齐机器人手形，提高接管和接触丰富示范质量。", "abstractZh": "DITTO-X同时提供关节力反馈与触点振动，并以适配映射支持Sharpa、Wuji和Inspire三种手。其反向遥操作让自主策略先驱动操作员手指至机器人当前配置，再交接控制，减少持物状态突变。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "real", "experimentNote": "6人用户研究比较Manus Pro，测试无视觉大小/软硬辨识、工具操作与策略接管。Franka Panda搭Sharpa执行夹钳、覆盆子、电池任务，固定示范80/60/60条；两轮DAgger各20干预轨迹，每策略30回合，并加等数据量非迭代对照。", "robots": [ "Franka Panda（实机）", "Sharpa Wave灵巧手（实机）", "Wuji 2灵巧手（实机）", "Inspire RH56DFX灵巧手（实机）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现需要手部运动学、力/电流标定、滤波与接管阈值，并保留限位、低于约0.08牛米电流限幅及急停。应使用相同起始条件和逐阶段成功规则，区分用户表现、示范质量与自主策略最终成功。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "三指主动外骨骼对食指/中指按关节映射，拇指经任务空间重定向，环指/小指跟随中指。利用电流或触觉估计载荷并回映射为关节反馈，Sharpa额外提供触点振动；反向模式投影机器人命令到人手，键盘请求与主动偏离阈值共同触发接管。", "whyUseful": "复现需要手部运动学、力/电流标定、滤波与接管阈值，并保留限位、低于约0.08牛米电流限幅及急停。应使用相同起始条件和逐阶段成功规则，区分用户表现、示范质量与自主策略最终成功。", "limitations": "六人样本小，主DAgger只在Sharpa完成，各接口修正自己的失败分布，不能完全隔离所有因果因素。三独立手指加耦合限制动作范围，振动只适用于有触觉传感的手；另两手反向模式未做DAgger评测。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00781v1", "note": "表I、§III，PDF页3–4：三手分别22/20/6驱动自由度；只有Sharpa支持触点振动。" }, { "url": "https://arxiv.org/pdf/2610.00781v1", "note": "§IV，PDF页5–6：Franka Panda、6人研究、30回合评测及等数量DAgger对照。" }, { "url": "https://arxiv.org/pdf/2610.00781v1", "note": "表II及§V，PDF页7–8：实机策略阶段成功率；Wuji/Inspire尚无DAgger结果。" }, { "url": "https://arxiv.org/pdf/2610.00781v1", "note": "PDF页7（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823245+00:00", "original": { "id": "arxiv-2610.00781", "originalSourceUrl": "https://arxiv.org/pdf/2610.00781v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00781v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00781", "sha256": "d1c4396096690834fbfc144e0e64ee870efd2de76ed28f6e11f471bac2271977", "bytes": 5584834, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:47.787358+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00781", "license": "http://creativecommons.org/licenses/by-nc-nd/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00781", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00781", "pages": 9, "sourceIdentityStatus": "title_match", "sourceTitle": "DITTO-X: Forward and Reverse Teleoperation for Dexterous Manipulation and Human Intervention", "sourceVersion": "2610.00781v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.659351+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00781" }, "originalAnalysis": { "id": "arxiv-2610.00781", "sectionsRead": [ "§III及表I，PDF页2–4", "§IV实验，PDF页5–6", "§V及图7–8、表II，PDF页6–8", "§VI，PDF页8" ], "analyzedAt": "2026-10-04T13:58:28.823245+00:00", "analysisStatus": "full_text_sections", "methodsZh": "三指主动外骨骼对食指/中指按关节映射，拇指经任务空间重定向，环指/小指跟随中指。利用电流或触觉估计载荷并回映射为关节反馈，Sharpa额外提供触点振动；反向模式投影机器人命令到人手，键盘请求与主动偏离阈值共同触发接管。", "experimentsZh": "6人用户研究比较Manus Pro，测试无视觉大小/软硬辨识、工具操作与策略接管。Franka Panda搭Sharpa执行夹钳、覆盆子、电池任务，固定示范80/60/60条；两轮DAgger各20干预轨迹，每策略30回合，并加等数据量非迭代对照。", "resultsZh": "完整反馈大小/软硬辨识86.1%/91.7%。示范任务成功70%对46.7%，接管78.3%对27.7%；DITTO-X第二轮DAgger三任务达86.7%、93.3%、90%。Wuji与Inspire另有示范策略验证。", "limitationsZh": "六人样本小，主DAgger只在Sharpa完成，各接口修正自己的失败分布，不能完全隔离所有因果因素。三独立手指加耦合限制动作范围，振动只适用于有触觉传感的手；另两手反向模式未做DAgger评测。", "reproductionZh": "复现需要手部运动学、力/电流标定、滤波与接管阈值，并保留限位、低于约0.08牛米电流限幅及急停。应使用相同起始条件和逐阶段成功规则，区分用户表现、示范质量与自主策略最终成功。", "sourceUrl": "https://arxiv.org/pdf/2610.00781v1", "sourceVersion": "2610.00781v1", "originalSha256": "d1c4396096690834fbfc144e0e64ee870efd2de76ed28f6e11f471bac2271977", "experimentType": "real", "robots": [ "Franka Panda（实机）", "Sharpa Wave灵巧手（实机）", "Wuji 2灵巧手（实机）", "Inspire RH56DFX灵巧手（实机）" ], "evidenceNotes": [ { "section": "表I、§III，PDF页3–4", "note": "三手分别22/20/6驱动自由度；只有Sharpa支持触点振动。" }, { "section": "§IV，PDF页5–6", "note": "Franka Panda、6人研究、30回合评测及等数量DAgger对照。" }, { "section": "表II及§V，PDF页7–8", "note": "实机策略阶段成功率；Wuji/Inspire尚无DAgger结果。" }, { "section": "PDF页7（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "高质量操作示范", "人机控制权交接" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Teleoperated demonstrations are a primary source of data for robot manipulation, and teleoperated interventions are a primary mechanism for correcting policies at deployment. Yet most teleoperation systems close the loop through vision alone and are built around parallel-jaw grippers, limiting both what the robot can execute and what the operator can express through it. This is most damaging in shared autonomy, where the operator sees the scene only through occluded cameras and must take over a dexterous hand mid-task, often with an object already grasped. We present DITTO-X, a hand-agnostic dexterous teleoperation interface that renders joint-level force and fingertip contact events from sensing already on the robot hand, and drives three commercial dexterous hands (Sharpa, Wuji, and Inspire) without per-hand redesign. Because the exoskeleton is actuated, DITTO-X also supports reverse teleoperation, in which the robot back-drives the operator's fingers into its own configuration before control is transferred, so the human enters the loop already matched to the state they inherit. Our results show that DITTO-X improves demonstration quality and throughput over a commercial hand-tracking glove, both in regular data collection and in human intervention during policy deployment for contact-rich manipulation tasks. More information can be found from our website: https://tml.stanford.edu/ditto-x/." }, { "id": "arxiv-2610.00801", "title": "ECoMEM: Explicit Concept Memory for Memory-Dependent Robot Control", "titleZh": "ECoMEM：面向依赖记忆的机器人控制的显式概念记忆", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00801", "paperUrl": "https://arxiv.org/abs/2610.00801", "projectUrl": null, "codeUrl": null, "category": "机器人记忆与长程操作", "tags": [ "显式记忆", "神经符号", "VLA", "事件计数" ], "directions": [ "视觉语言动作" ], "robotFilters": [ "YAM" ], "tier": "recent", "summary": "将证据支持的记忆写入与动作学习分开，显著改善遮挡、计数及顺序任务。", "abstractZh": "ECoMEM以可复用概念库记录实体、关系、事件进展和时间顺序。固定Writer按视觉与机器人状态维护记录，学习型Reader将其变为VLA记忆token。原文在16项RoboMME仿真任务和两项YAM实体任务验证复用与扩展。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "both", "experimentNote": "RoboMME共16任务、20万更新，每任务50回合及3策略采样种子，基线使用基准论文已报结果。实机YAM右臂进行豆子舀倒和杯子往返，每任务59条共同示范；前者增加SCOOP概念，后者复用原库。", "robots": [ "YAM右臂（实机）", "RoboMME任务机械臂（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现关键是只使用当前以前证据、冻结Writer、随机置换实体ID、防止示范事件误算为执行进度。需报告概念开发代价、记忆初始化与更新延迟；YAM部署20步预测执行16步，记忆更新5/10Hz。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "Writer先按指令选择概念和依赖，用OWLv2定位、SAM2跟踪及末端/夹爪状态建立证据，经过确认、去重、计数与前置条件规则更新记忆。Reader对11类离散字段嵌入，以小Transformer逐记录生成token，与π0.5语言主干和动作专家联合训练。", "whyUseful": "复现关键是只使用当前以前证据、冻结Writer、随机置换实体ID、防止示范事件误算为执行进度。需报告概念开发代价、记忆初始化与更新延迟；YAM部署20步预测执行16步，记忆更新5/10Hz。", "limitations": "实机试验非配对且数量不等，只有两种重复操作。概念及识别规则经人工/编码代理设计验证，不是自动学得通用记忆；感知错绑无法由策略可靠修复。插孔仍仅40.67%，正确记忆不能替代精确控制。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00801v1", "note": "§4.2–4.3，PDF页6–7：概念逻辑显式构建、Writer固定，Reader与VLA训练。" }, { "url": "https://arxiv.org/pdf/2610.00801v1", "note": "表1，PDF页8；§5.5，PDF页12：仿真基线来自RoboMME公开报告，实机68/79与5/58非配对不等量。" }, { "url": "https://arxiv.org/pdf/2610.00801v1", "note": "附录I，PDF页27–28：YAM右臂每任务59条示范，6关节加夹爪状态，记忆在线因果更新。" }, { "url": "https://arxiv.org/pdf/2610.00801v1", "note": "PDF页11（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823243+00:00", "original": { "id": "arxiv-2610.00801", "originalSourceUrl": "https://arxiv.org/pdf/2610.00801v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00801v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00801", "sha256": "1a1b71465841402967dfd31707375269b9c185a6a821fbccac3d10ed86934f3a", "bytes": 9099114, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:35.349937+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00801", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00801", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00801", "pages": 29, "sourceIdentityStatus": "title_match", "sourceTitle": "ECoMEM: Explicit Concept Memory for Memory-Dependent Robot Control", "sourceVersion": "2610.00801v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.652374+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00801" }, "originalAnalysis": { "id": "arxiv-2610.00801", "sectionsRead": [ "§4，PDF页5–7", "§5.1–5.4及表1–2，PDF页8–11", "§5.5及结论，PDF页12", "附录I与表8–9，PDF页27–29" ], "analyzedAt": "2026-10-04T13:58:28.823243+00:00", "analysisStatus": "full_text_sections", "methodsZh": "Writer先按指令选择概念和依赖，用OWLv2定位、SAM2跟踪及末端/夹爪状态建立证据，经过确认、去重、计数与前置条件规则更新记忆。Reader对11类离散字段嵌入，以小Transformer逐记录生成token，与π0.5语言主干和动作专家联合训练。", "experimentsZh": "RoboMME共16任务、20万更新，每任务50回合及3策略采样种子，基线使用基准论文已报结果。实机YAM右臂进行豆子舀倒和杯子往返，每任务59条共同示范；前者增加SCOOP概念，后者复用原库。", "resultsZh": "仿真平均82.42%，最强列举基线44.51%，15/16任务领先。实机按恰好完成指定次数判定，ECoMEM为68/79（86.1%），无记忆π0.5为5/58（8.6%）；去除空间、事件或顺序信息均明显恶化对应任务。", "limitationsZh": "实机试验非配对且数量不等，只有两种重复操作。概念及识别规则经人工/编码代理设计验证，不是自动学得通用记忆；感知错绑无法由策略可靠修复。插孔仍仅40.67%，正确记忆不能替代精确控制。", "reproductionZh": "复现关键是只使用当前以前证据、冻结Writer、随机置换实体ID、防止示范事件误算为执行进度。需报告概念开发代价、记忆初始化与更新延迟；YAM部署20步预测执行16步，记忆更新5/10Hz。", "sourceUrl": "https://arxiv.org/pdf/2610.00801v1", "sourceVersion": "2610.00801v1", "originalSha256": "1a1b71465841402967dfd31707375269b9c185a6a821fbccac3d10ed86934f3a", "experimentType": "both", "robots": [ "YAM右臂（实机）", "RoboMME任务机械臂（仿真）" ], "evidenceNotes": [ { "section": "§4.2–4.3，PDF页6–7", "note": "概念逻辑显式构建、Writer固定，Reader与VLA训练。" }, { "section": "表1，PDF页8；§5.5，PDF页12", "note": "仿真基线来自RoboMME公开报告，实机68/79与5/58非配对不等量。" }, { "section": "附录I，PDF页27–28", "note": "YAM右臂每任务59条示范，6关节加夹爪状态，记忆在线因果更新。" }, { "section": "PDF页11（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "结构化机器人记忆", "长程任务状态追踪" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "A robot may lose sight of an object it must later retrieve, need to recall what a person demonstrated earlier, or track which steps of a task it has already completed. Current vision-language-action (VLA) policies often fail once the information needed for action disappears from the current observation, making memory critical for long-horizon robot behavior. Existing approaches typically provide longer histories or learn implicit memory from observation-action trajectories. But action supervision tells a policy how to act, not what to remember: it does not specify which past facts should persist or how they should change as new evidence arrives. We therefore separate maintaining an evidence-grounded account of the past from learning how to act on it. This insight motivates Explicit Concept Memory (ECoMEM), which represents task-relevant history with a reusable library of grounded concepts. An evidence-based Writer selects and updates these records, while a learned Reader turns them into memory tokens that directly condition the VLA. Across 16 RoboMME tasks, ECoMEM leads the evaluated robot policies on 15 tasks. On two new real-robot tasks, the same memory library either transfers directly or requires only one new concept, achieving 86.1% success versus 8.6% for a no-memory VLA. These results show that explicit concepts provide a reusable and extensible memory interface for robot control. Project website: https://ecomem.github.io/" }, { "id": "arxiv-2610.00822", "title": "TRACE: Privacy-Preserving Next-Best-View Selection over Distributed 3D Gaussian-Splat Maps", "titleZh": "TRACE：在分布式三维高斯泼溅地图上保护隐私的下一最佳视角选择", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00822", "paperUrl": "https://arxiv.org/abs/2610.00822", "projectUrl": null, "codeUrl": null, "category": "多机器人感知与建图", "tags": [ "3DGS", "分布式主动感知", "下一最佳视角", "地图隐私" ], "directions": [ "导航与建图" ], "robotFilters": [], "tier": "recent", "summary": "只交换射线深度聚合量，近似集中式地图的视角信息增益与梯度。", "abstractZh": "TRACE让各机器人保留自己的高斯地图，仅发送透射率、背景辐射相关深度统计及姿态导数。理论分析给出无混合深度分箱时的精确重建条件；Habitat-Sim实验显示接近集中式视角收益，但稠密通信代价很高。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "sim", "experimentNote": "Habitat-Sim和Gibson场景测试导航/探索，比较独立规划、集中地图共享、集中oracle及TRACE稀疏版本。四机器人跨3种子进行文中所称100个视角决策，优化方向再投到24个可执行朝向；无实体机器人实验。", "robots": [ "Habitat-Sim RGB-D移动代理（仿真）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现需相同相机位姿、径向深度网格、掩码路径及局部gsplat地图，并保留固定命中集合等理论前提。除PSNR/SSIM外应测含传输的端到端时延、每步团队载荷及分箱精度，核查隐私威胁模型。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "证明跨地图耦合可归约为高斯前方透射率与后方辐射，两者按射线深度箱分布式求和。各机器人在路径掩码范围交换统计和旋转导数，重建信息增益并在SO(3)优化朝向；无混合分箱时精确，否则由光学深度给出误差界。", "whyUseful": "复现需相同相机位姿、径向深度网格、掩码路径及局部gsplat地图，并保留固定命中集合等理论前提。除PSNR/SSIM外应测含传输的端到端时延、每步团队载荷及分箱精度，核查隐私威胁模型。", "limitations": "地图参数不共享不等于零信息泄露：原文说明统计会暴露粗密度和彩色体积，未提供差分隐私保证。时间统计排除网络延迟，稠密通信可能主导部署；质量在各方法实际走过轨迹上评测，非完全相同视点。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00822v1", "note": "§VI-B，PDF页5：明确聚合消息仍披露像素/深度箱分辨率的粗密度与颜色。" }, { "url": "https://arxiv.org/pdf/2610.00822v1", "note": "表II与说明，PDF页7：稠密载荷1004–2055MB/步；NBV时间不含网络延迟。" }, { "url": "https://arxiv.org/pdf/2610.00822v1", "note": "§VII-E表IV，PDF页8：100决策、3种子表述及97.9%EIG；没有实机部署。" }, { "url": "https://arxiv.org/pdf/2610.00822v1", "note": "PDF页7（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823241+00:00", "original": { "id": "arxiv-2610.00822", "originalSourceUrl": "https://arxiv.org/pdf/2610.00822v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00822v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00822", "sha256": "a16b686f990ad1e024a8af24884931d0e1021e3f04e5ef32f144923264391fda", "bytes": 1144846, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:35.099317+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00822", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00822", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00822", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "TRACE: Privacy-Preserving Next-Best-View Selection over Distributed 3D Gaussian-Splat Maps", "sourceVersion": "2610.00822v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.642586+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00822" }, "originalAnalysis": { "id": "arxiv-2610.00822", "sectionsRead": [ "§V–VI，PDF页4–6", "§VII-A–D及表I–III，PDF页6–8", "§VII-E表IV及结论，PDF页8" ], "analyzedAt": "2026-10-04T13:58:28.823241+00:00", "analysisStatus": "full_text_sections", "methodsZh": "证明跨地图耦合可归约为高斯前方透射率与后方辐射，两者按射线深度箱分布式求和。各机器人在路径掩码范围交换统计和旋转导数，重建信息增益并在SO(3)优化朝向；无混合分箱时精确，否则由光学深度给出误差界。", "experimentsZh": "Habitat-Sim和Gibson场景测试导航/探索，比较独立规划、集中地图共享、集中oracle及TRACE稀疏版本。四机器人跨3种子进行文中所称100个视角决策，优化方向再投到24个可执行朝向；无实体机器人实验。", "resultsZh": "TRACE与集中式朝向完全匹配77.8%、15度内83.3%，保留97.9%信息增益。三至十机器人稠密通信约1004–2055MB/步，稀疏版1.45–8.42MB/步，后者以部分重建质量换通信下降。", "limitationsZh": "地图参数不共享不等于零信息泄露：原文说明统计会暴露粗密度和彩色体积，未提供差分隐私保证。时间统计排除网络延迟，稠密通信可能主导部署；质量在各方法实际走过轨迹上评测，非完全相同视点。", "reproductionZh": "复现需相同相机位姿、径向深度网格、掩码路径及局部gsplat地图，并保留固定命中集合等理论前提。除PSNR/SSIM外应测含传输的端到端时延、每步团队载荷及分箱精度，核查隐私威胁模型。", "sourceUrl": "https://arxiv.org/pdf/2610.00822v1", "sourceVersion": "2610.00822v1", "originalSha256": "a16b686f990ad1e024a8af24884931d0e1021e3f04e5ef32f144923264391fda", "experimentType": "sim", "robots": [ "Habitat-Sim RGB-D移动代理（仿真）" ], "evidenceNotes": [ { "section": "§VI-B，PDF页5", "note": "明确聚合消息仍披露像素/深度箱分辨率的粗密度与颜色。" }, { "section": "表II与说明，PDF页7", "note": "稠密载荷1004–2055MB/步；NBV时间不含网络延迟。" }, { "section": "§VII-E表IV，PDF页8", "note": "100决策、3种子表述及97.9%EIG；没有实机部署。" }, { "section": "PDF页7（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "跨机器人信息增益", "通信与重建权衡" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "Share the light, not the map. We study next-best-view selection for a team of robots, each of which builds its own 3D Gaussian Splatting map and keeps it private. A robot picks the view with the largest expected information gain (EIG) about the splats along its own path. This gain depends on the other maps. Their splats occlude its own and shine behind them, so the gain has to be evaluated against the pooled map. No robot has this map. We show that the coupling passes through only two ray quantities, the transmittance in front of a splat and the radiance behind it, and that both are sums over the hits of the ray. Hence, they decompose across the robots, and each robot sums them over depth bins in its own map, along the rays of a candidate view, and sends the sums with their pose derivatives. The robot planning the view turns them into its EIG and gradient on SO(3). Transmittance and Radiance Aggregates, communicated for the EIG, give the protocol its name: TRACE. No robot shares its splats, and the message size does not grow with a map. We prove that the reconstruction is exact unless a depth bin behind a splat mixes hits of two robots, and we bound the error otherwise. Over 100 next-best-view decisions in Habitat-Sim, TRACE picks a heading within 15 degrees of the centralized one in 83.3% of the cases, and its views reach 97.9% of the centralized EIG." }, { "id": "arxiv-2610.00823", "title": "Reactive Humanoid Multi-Contact Using Learned Stability Models", "titleZh": "利用学习型稳定性模型实现人形机器人的反应式多接触控制", "date": "2026-09-30", "year": 2026, "url": "https://arxiv.org/abs/2610.00823", "paperUrl": "https://arxiv.org/abs/2610.00823", "projectUrl": null, "codeUrl": null, "category": "人形机器人控制", "tags": [ "抗扰恢复", "多接触", "稳定性学习", "支撑规划" ], "directions": [ "运动控制" ], "robotFilters": [ "Alex" ], "tier": "recent", "summary": "用神经网络近似多接触稳定区域，让人形机器人在受推后迅速选取手部支撑。", "abstractZh": "系统先学习手脚接触配置对应的可行压力中心区域，再预测受扰后的质心运动，选择支撑平面与接触点。Alex人形机器人的仿真和硬件推扰试验显示，利用手部支撑可增加可承受冲量并缩短恢复时间。", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "experimentType": "both", "experimentNote": "Alex配头部ZED X Mini提取30Hz平面图。仿真在站立、侧行、后退中二分搜索最大可承受冲量；硬件做双表面站立、单墙站立各10次推扰及行走12次，对比最近可达支撑或无手支撑。", "robots": [ "Alex人形机器人（仿真与实机）" ], "robotNote": "型号和使用场景以原文及上列说明为准；训练或离线采集平台不等于所提方法的实机闭环验证。", "codeStatus": "unknown", "status": "代码状态未核实", "codeStatusNote": "本次核对论文原文，未独立验证代码仓库内容、权重、训练入口与代码许可证。", "trainingNote": "复现需机器人惯量、关节力矩/摩擦约束与各接触训练集：单手5万、双手10万样本。需重现安全缩边7厘米、触发阈值、可达域及接触检测；不能把该平台网络直接当跨人形通用稳定性模型。 论文许可不代表代码许可，代码实际发布状态仍未知。", "contribution": "将满足摩擦、单向接触及关节力矩约束的线性规划结果作为监督，对5类镜像去重的手脚配置各训网络，输出18方向上的压力中心扩展。候选支撑先按平面再按点筛选，滚动预测接触前、碰撞及接触后捕获点，用可恢复控制余量评分并接入逆动力学。", "whyUseful": "复现需机器人惯量、关节力矩/摩擦约束与各接触训练集：单手5万、双手10万样本。需重现安全缩边7厘米、触发阈值、可达域及接触检测；不能把该平台网络直接当跨人形通用稳定性模型。", "limitations": "足部假设近似平面，左右手独立规划且简化碰撞冲量。实机未直接测力，以捕获点斜率近似推力，行走基线为未触发支撑的推扰，严格配对有限；手释放后的动量和检测延迟仍影响恢复。", "caveats": "性能数字为作者报告，未独立复现实验；阅读范围为列出的原文关键章节。", "evidence": [ { "url": "https://arxiv.org/pdf/2610.00823v1", "note": "§III-B，PDF页3：5类配置、18方向输出、优化生成训练数据。" }, { "url": "https://arxiv.org/pdf/2610.00823v1", "note": "表II–III与§VI-C，PDF页6–7：实机推力未直接测量；行走比较协议与站立不同。" }, { "url": "https://arxiv.org/pdf/2610.00823v1", "note": "PDF页7（已渲染目视核对）：核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "verificationNote": "核对原文身份、所列章节、实验类别与明确型号；不代表读完或翻译了每一页。", "verifiedAt": "2026-10-04", "analysisVerifiedAt": "2026-10-04T13:58:28.823239+00:00", "original": { "id": "arxiv-2610.00823", "originalSourceUrl": "https://arxiv.org/pdf/2610.00823v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2610.00823v1", "finalDownloadUrl": "https://arxiv.org/pdf/2610.00823", "sha256": "c178065b037923e2cebc7d02c2ae4b98fcbbef077a488fd2b2d3e922f9d5b6b0", "bytes": 4810163, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:43:30.163901+00:00", "publicationDate": "2026/09/30", "publicationDateSourceUrl": "https://arxiv.org/abs/2610.00823", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2610.00823", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2610.00823", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "Reactive Humanoid Multi-Contact Using Learned Stability Models", "sourceVersion": "2610.00823v1", "sourceVersionDate": "2026/09/30", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.638701+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2610.00823" }, "originalAnalysis": { "id": "arxiv-2610.00823", "sectionsRead": [ "§III，PDF页2–3", "§IV–V，PDF页4–6", "§VI–VII及表I–III，PDF页6–7" ], "analyzedAt": "2026-10-04T13:58:28.823239+00:00", "analysisStatus": "full_text_sections", "methodsZh": "将满足摩擦、单向接触及关节力矩约束的线性规划结果作为监督，对5类镜像去重的手脚配置各训网络，输出18方向上的压力中心扩展。候选支撑先按平面再按点筛选，滚动预测接触前、碰撞及接触后捕获点，用可恢复控制余量评分并接入逆动力学。", "experimentsZh": "Alex配头部ZED X Mini提取30Hz平面图。仿真在站立、侧行、后退中二分搜索最大可承受冲量；硬件做双表面站立、单墙站立各10次推扰及行走12次，对比最近可达支撑或无手支撑。", "resultsZh": "仿真最大冲量分别29.4/42.3/32.1牛顿秒，较无支撑平均增89%、最近点增17%。实机站立两条件恢复时间减少57%和29%，行走从384降至316毫秒。规划总时间9.7毫秒，数值方法388毫秒。", "limitationsZh": "足部假设近似平面，左右手独立规划且简化碰撞冲量。实机未直接测力，以捕获点斜率近似推力，行走基线为未触发支撑的推扰，严格配对有限；手释放后的动量和检测延迟仍影响恢复。", "reproductionZh": "复现需机器人惯量、关节力矩/摩擦约束与各接触训练集：单手5万、双手10万样本。需重现安全缩边7厘米、触发阈值、可达域及接触检测；不能把该平台网络直接当跨人形通用稳定性模型。", "sourceUrl": "https://arxiv.org/pdf/2610.00823v1", "sourceVersion": "2610.00823v1", "originalSha256": "c178065b037923e2cebc7d02c2ae4b98fcbbef077a488fd2b2d3e922f9d5b6b0", "experimentType": "both", "robots": [ "Alex人形机器人（仿真与实机）" ], "evidenceNotes": [ { "section": "§III-B，PDF页3", "note": "5类配置、18方向输出、优化生成训练数据。" }, { "section": "表II–III与§VI-C，PDF页6–7", "note": "实机推力未直接测量；行走比较协议与站立不同。" }, { "section": "PDF页7（已渲染目视核对）", "note": "核对该页关键图表的行列对应、实验类别与数值；分析范围仅限sectionsRead所列章节，并非逐页翻译。" } ], "verification": "verified", "directionsOriginal": [ "反应式手部支撑", "人形平衡恢复" ] }, "firstSeen": "2026-10-04T12:29:36.801Z", "lastSeen": "2026-10-04T12:30:42.175Z", "abstractOriginal": "We present a planning and control approach to reactively use hand contacts to stabilize a humanoid in low stability scenarios, where only using feet contacts may result in a fall. Candidate contacts are sampled within the robot's reachable workspace, and a preview is computed by rolling out the centroidal dynamics through pre-impact, impact and post-impact phases. Sampled points are scored based on the Center of Pressure (CoP) control authority at the post-impact phase. Central to our approach is a learned model of the robot's CoP region during post-impact, which enables rapid evaluation of candidate contact points compared to traditional optimization-based methods. The presented planner has two stages: the first selects an optimal bracing region and the second computes an optimal bracing point within the region. Our simulation results demonstrate an average increase in impulse resilience of 89% over recovery without hand contacts and 17% over a naive planning strategy (closest reachable region). We validate our framework on hardware, performing push tests while standing and walking. The standing trials show an average 43% reduction in stabilization time compared to naive hand placement and the walking trials demonstrate a 18% reduction compared to baseline recovery (without hand contacts)." }, { "id": "arxiv-2609.34018", "title": "Estimate, Don't Imitate: Reusing Differentiable State-Based Policies for Visuomotor Control", "shortTitle": "Estimate, Don't Imitate", "date": "2026-09-27", "category": "机器人学习 / Sim-to-Real", "freshness": "最新预印本", "summary": "保留已训练且可微分的状态控制专家，只学习视觉到其缺失物理状态的估计器。多相机四帧视觉和本体状态经DINOv3-LoRA、时序卷积输出物体及目标相关状态，以状态回归和穿过冻结专家的动作一致性混合监督，逐渐增加动作…", "whyUseful": "需先得到可微冻结专家及其准确状态定义，再固定状态标准化、损失渐变和训练预算。完整报告安全中止与重启，区分Panda实机和Allegro仿真；未独立复现。", "paperUrl": "https://arxiv.org/abs/2609.34018", "projectUrl": null, "codeUrl": null, "trainingStatus": "未核实公开训练代码", "trainingNote": "已检查论文全文和相关搜索，未核实作者训练仓库。", "caveats": "76% 是 PandaCube 实机 19/25 的作者结果，五类任务是仿真评估；若把两次安全中止计为失败，实机为 70.4%。不能外推为通用成功率。", "url": "https://arxiv.org/abs/2609.34018", "status": "未核实公开训练代码", "titleZh": "估计，而非模仿：复用可微的状态策略实现视觉运动控制", "abstractZh": "保留已训练且可微分的状态控制专家，只学习视觉到其缺失物理状态的估计器。多相机四帧视觉和本体状态经DINOv3-LoRA、时序卷积输出物体及目标相关状态，以状态回归和穿过冻结专家的动作一致性混合监督，逐渐增加动作损失。\n在五种MuJoCo操作任务比较状态估计、像素行为克隆和完整状态克隆；相同4万步训练预算，主要条件三种子、每次1000闭环回合。真实Panda仅测PandaCube，以视觉域随机化仿真训练，固定初始位置并改变方块偏航。", "experimentType": "both", "experimentNote": "在五种MuJoCo操作任务比较状态估计、像素行为克隆和完整状态克隆；相同4万步训练预算，主要条件三种子、每次1000闭环回合。真实Panda仅测PandaCube，以视觉域随机化仿真训练，固定初始位置并改变方块偏航。", "robots": [ "Franka Panda", "Allegro Hand（仿真）" ], "robotNote": "论文实机称 Panda；Allegro 为仿真中的 16 自由度灵巧手，未核实更细子型号。", "tags": [ "状态估计", "Sim-to-Real", "可微控制", "特权专家", "接触操作" ], "contribution": "保留已训练且可微分的状态控制专家，只学习视觉到其缺失物理状态的估计器。多相机四帧视觉和本体状态经DINOv3-LoRA、时序卷积输出物体及目标相关状态，以状态回归和穿过冻结专家的动作一致性混合监督，逐渐增加动作损失。", "limitations": "不同任务最优权重不同，单纯动作损失可能崩溃；部分克隆对照仅单种子。优势是在匹配预算下而非所有方法训练收敛后的结论。真实测试规模小，仅一任务，接触动力学失配仍导致失败。", "codeStatus": "unknown", "codeStatusNote": "未核实论文对应的官方公开实现，不能据此断言作者不提供代码。", "license": "未核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/html/2609.34018v1", "note": "摘要、§IV-A、§IV-D、§V-D 及表 IV：方法、仿真任务、PandaCube 实机及试验分母；附录说明匹配训练预算。" }, { "url": "https://arxiv.org/pdf/2609.34018", "note": "§III; Table II：冻结专家、状态估计与任务依赖的混合损失效果。" }, { "url": "https://arxiv.org/pdf/2609.34018", "note": "Table IV; Appendix A：25/27尝试分母差别；克隆延长预算的控制实验。" } ], "tier": "recent", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对范围为公开原文、平台型号与所列来源；性能数字均为作者报告，未独立复现。", "year": 2026, "directions": [ "操作与抓取", "模仿学习" ], "robotFilters": [ "Franka Panda", "Allegro Hand" ], "original": { "id": "arxiv-2609.34018", "originalSourceUrl": "https://arxiv.org/pdf/2609.34018v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.34018v1", "finalDownloadUrl": "https://arxiv.org/pdf/2609.34018", "sha256": "de94b28080ee3f4a10971a551b7cee14ea976b7c0066644663b523c6c7b25a2f", "bytes": 3405417, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:33:10.712064+00:00", "publicationDate": "2026/09/27", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.34018", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.34018", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.34018", "pages": 9, "sourceIdentityStatus": "title_match", "sourceTitle": "Estimate, Don't Imitate: Reusing Differentiable State-Based Policies for Visuomotor Control", "sourceVersion": "2609.34018v1", "sourceVersionDate": "2026/09/27", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.595046+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.34018" }, "originalAnalysis": { "id": "arxiv-2609.34018", "sourceUrl": "https://arxiv.org/pdf/2609.34018", "sectionsRead": [ "§III", "§IV–V", "Tables II–IV", "Appendix A" ], "methodsZh": "保留已训练且可微分的状态控制专家，只学习视觉到其缺失物理状态的估计器。多相机四帧视觉和本体状态经DINOv3-LoRA、时序卷积输出物体及目标相关状态，以状态回归和穿过冻结专家的动作一致性混合监督，逐渐增加动作损失。", "experimentsZh": "在五种MuJoCo操作任务比较状态估计、像素行为克隆和完整状态克隆；相同4万步训练预算，主要条件三种子、每次1000闭环回合。真实Panda仅测PandaCube，以视觉域随机化仿真训练，固定初始位置并改变方块偏航。", "resultsZh": "作者报告混合损失在PandaCube达99.2%，TrayPlate从仅状态监督28.9%升至64.1%；AllegroCube反而77.2%低于仅状态监督79.2%。实机25个完成评测回合成功76%，若把另两次关节限位安全中止计作失败，则19/27为70.4%。", "limitationsZh": "不同任务最优权重不同，单纯动作损失可能崩溃；部分克隆对照仅单种子。优势是在匹配预算下而非所有方法训练收敛后的结论。真实测试规模小，仅一任务，接触动力学失配仍导致失败。", "reproductionZh": "需先得到可微冻结专家及其准确状态定义，再固定状态标准化、损失渐变和训练预算。完整报告安全中止与重启，区分Panda实机和Allegro仿真；未独立复现。", "experimentType": "both", "robots": [ "Franka Panda", "Allegro Hand（仿真）" ], "evidenceNotes": [ { "section/page": "§III; Table II", "note": "冻结专家、状态估计与任务依赖的混合损失效果。" }, { "section/page": "Table IV; Appendix A", "note": "25/27尝试分母差别；克隆延长预算的控制实验。" } ], "corrections": [ { "field": "resultsNote", "value": "实机76%排除两次安全中止；所有27次尝试计入为70.4%", "reason": "原文Table IV脚注明确给出两种分母。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04" }, "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2609.28959", "title": "TactileStep: Sole Tactile Learning for Regulating Foot-Terrain Interaction in Humanoid Locomotion", "shortTitle": "TactileStep", "date": "2026-09-24", "category": "人形机器人 / 触觉行走", "freshness": "最新论文", "summary": "把足底法向力、压力中心和接触面积压缩成可部署特征，与本体和深度历史共同输入PPO策略。轻量触觉模拟将刚体合力按60个触元间隙与法向分配，再模拟空间扩散；按步态阶段奖励柔和落地及稳固支撑，双critic分担稠密和…", "whyUseful": "优先复现触元标定与仿真/实机特征一致性，再分离在线触觉和奖励塑形作用。记录力采样与首个上升沿冲击定义；同步报告接触质量、速度、能耗，防止只用通行成功掩盖接触差异。", "paperUrl": "https://arxiv.org/abs/2609.28959", "projectUrl": "https://tactilestep.github.io/", "codeUrl": null, "trainingStatus": "待开源", "trainingNote": "官方项目页标注 Code Soon。", "caveats": "项目页标注 CoRL 2026 Spotlight；需额外足底传感硬件，尚未核实可训练仓库。", "url": "https://arxiv.org/abs/2609.28959", "status": "待开源", "titleZh": "TactileStep：通过足底触觉学习调节人形机器人行走中的足部—地形交互", "abstractZh": "把足底法向力、压力中心和接触面积压缩成可部署特征，与本体和深度历史共同输入PPO策略。轻量触觉模拟将刚体合力按60个触元间隙与法向分配，再模拟空间扩散；按步态阶段奖励柔和落地及稳固支撑，双critic分担稠密和事件式信号。\n29自由度Unitree G1在Isaac Sim/Isaac Lab训练，2048并行；仿真每策略每地形4096回合，实机每条件20样本。比较视觉跑酷基线、无触觉观察、无软落地及无稳定支撑，在楼梯、平台、斜坡和平地测力、面积与声压。", "experimentType": "both", "experimentNote": "29自由度Unitree G1在Isaac Sim/Isaac Lab训练，2048并行；仿真每策略每地形4096回合，实机每条件20样本。比较视觉跑酷基线、无触觉观察、无软落地及无稳定支撑，在楼梯、平台、斜坡和平地测力、面积与声压。", "robots": [ "Unitree G1" ], "robotNote": "加装足底压力鞋垫；不是裸机默认传感配置。", "tags": [ "人形机器人", "足底触觉", "强化学习", "行走控制", "接触质量" ], "contribution": "把足底法向力、压力中心和接触面积压缩成可部署特征，与本体和深度历史共同输入PPO策略。轻量触觉模拟将刚体合力按60个触元间隙与法向分配，再模拟空间扩散；按步态阶段奖励柔和落地及稳固支撑，双critic分担稠密和事件式信号。", "limitations": "柔和接触有能耗代价，不能声称全面效率提升；声学受材料及约55–60dB风扇底噪影响，应地形内比较。高速、软土颗粒表面、鞋垫长期漂移耐久及系统性失败恢复未验证。", "codeStatus": "pending", "codeStatusNote": "官方项目页明确标注 Code Soon。", "license": "未核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/html/2609.28959v1", "note": "摘要、§4 及 §6 支持实验类型、G1 型号、指标和局限。" }, { "url": "https://tactilestep.github.io/", "note": "官方项目页展示仿真/实机实验并标注 Code Soon。" }, { "url": "https://arxiv.org/pdf/2609.28959", "note": "3.1–3.2; 4.1：60触元近似、actor可部署输入及重复数量。" }, { "url": "https://arxiv.org/pdf/2609.28959", "note": "Tables 2–3; 6：硬件量化、能耗代价和未验证范围。" }, { "url": "https://arxiv.org/pdf/2609.28959v1", "note": "3.1–3.2; 4.1：60触元近似、actor可部署输入及重复数量。" }, { "url": "https://arxiv.org/pdf/2609.28959v1", "note": "Tables 2–3; 6：硬件量化、能耗代价和未验证范围。" } ], "tier": "recent", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对范围为公开原文、平台型号与所列来源；性能数字均为作者报告，未独立复现。", "year": 2026, "directions": [ "运动控制", "人形机器人", "强化学习" ], "robotFilters": [ "Unitree G1" ], "original": { "id": "arxiv-2609.28959", "originalSourceUrl": "https://arxiv.org/pdf/2609.28959v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.28959v1", "finalDownloadUrl": "https://arxiv.org/pdf/2609.28959", "sha256": "d08d4cfbcd719d59aadf34325c7763c9b8c6b029e8c85435a5f652705c798426", "bytes": 5222638, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:33:10.696557+00:00", "publicationDate": "2026/09/24", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.28959", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.28959", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.28959", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "TactileStep: Sole Tactile Learning for Regulating Foot-Terrain Interaction in Humanoid Locomotion", "sourceVersion": "2609.28959v1", "sourceVersionDate": "2026/09/24", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.600053+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.28959" }, "originalAnalysis": { "id": "arxiv-2609.28959", "sourceUrl": "https://arxiv.org/pdf/2609.28959v1", "sectionsRead": [ "3.1–3.4", "4.1–4.3", "6", "Tables 1–3" ], "methodsZh": "把足底法向力、压力中心和接触面积压缩成可部署特征，与本体和深度历史共同输入PPO策略。轻量触觉模拟将刚体合力按60个触元间隙与法向分配，再模拟空间扩散；按步态阶段奖励柔和落地及稳固支撑，双critic分担稠密和事件式信号。", "experimentsZh": "29自由度Unitree G1在Isaac Sim/Isaac Lab训练，2048并行；仿真每策略每地形4096回合，实机每条件20样本。比较视觉跑酷基线、无触觉观察、无软落地及无稳定支撑，在楼梯、平台、斜坡和平地测力、面积与声压。", "resultsZh": "实机上平台冲击从695.0N降到355.7N，下楼接触面积0.483升到0.598；各地形噪声均下降。仿真成功率不低于基线，但速度跟踪误差和耗时一般略升，能耗更高，例如上楼1126.3J对802.2J。", "limitationsZh": "柔和接触有能耗代价，不能声称全面效率提升；声学受材料及约55–60dB风扇底噪影响，应地形内比较。高速、软土颗粒表面、鞋垫长期漂移耐久及系统性失败恢复未验证。", "reproductionZh": "优先复现触元标定与仿真/实机特征一致性，再分离在线触觉和奖励塑形作用。记录力采样与首个上升沿冲击定义；同步报告接触质量、速度、能耗，防止只用通行成功掩盖接触差异。", "experimentType": "both", "robots": [ "Unitree G1" ], "corrections": [ "降低冲击和噪声不等于降低能耗，表2反而报告更高能耗。" ], "evidenceNotes": [ { "note": "60触元近似、actor可部署输入及重复数量。", "section": "3.1–3.2; 4.1" }, { "note": "硬件量化、能耗代价和未验证范围。", "section": "Tables 2–3; 6" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2609.28959v1", "originalSha256": "d08d4cfbcd719d59aadf34325c7763c9b8c6b029e8c85435a5f652705c798426" }, "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2609.28258", "title": "Generalizable Robotic Insertion with World Models", "shortTitle": "InsertionWM", "date": "2026-09-23", "category": "世界模型 / 装配 / Model-based RL", "freshness": "近期 arXiv 版本；有 2025 前身", "summary": "InsertionWM以TD-MPC2学习潜在动力学、奖励、价值与策略先验，用MPPI在潜空间滚动规划。输入腕部96×96深度及24维本体/带噪位姿，输出六维末端增量；多装配共享模型，按各装配成功率调整初始状态…", "whyUseful": "复现需AutoMate资产、奖励和预抓取流程，注意训练仍用CAD构建仿真以及带噪插座位姿/目标位置。模型500万参数，通用训练约三天RTX3090；无部署CAD不等于整个训练流程无CAD，本轮未运行。", "paperUrl": "https://arxiv.org/abs/2609.28258", "projectUrl": null, "codeUrl": null, "trainingStatus": "未核实官方训练代码", "trainingNote": "基于 Isaac/AutoMate 仿真和 TD-MPC2；未核实该论文完整公开实现。", "caveats": "全文明确所有实验均在仿真，不能写成实机结果。作者 CV 列出同题 2025 RSS workshop 前身；9 月日期是 arXiv 上传日期。", "extraSources": [ "https://arxiv.org/html/2609.28258v1", "https://www.nicklashansen.com/files/cv.pdf", "https://openreview.net/pdf?id=DR3n6IqGKI" ], "url": "https://arxiv.org/abs/2609.28258", "status": "未核实官方训练代码", "titleZh": "利用世界模型实现可泛化的机器人插入装配", "abstractZh": "InsertionWM以TD-MPC2学习潜在动力学、奖励、价值与策略先验，用MPPI在潜空间滚动规划。输入腕部96×96深度及24维本体/带噪位姿，输出六维末端增量；多装配共享模型，按各装配成功率调整初始状态课程。\n100套双零件装配资产、仿真Franka Panda；90训练、按ID留出10套同分布未见几何。专用策略各测5种子、每次100回合，通用模型每资产100次；与AutoMate、无深度、不同视角和课程等比较。", "experimentType": "sim", "experimentNote": "100套双零件装配资产、仿真Franka Panda；90训练、按ID留出10套同分布未见几何。专用策略各测5种子、每次100回合，通用模型每资产100次；与AutoMate、无深度、不同视角和课程等比较。", "robots": [ "Franka Panda" ], "robotNote": "Isaac/AutoMate 内的仿真机器人，配平行夹爪和模拟腕部深度相机。", "tags": [ "世界模型", "装配", "模型式强化学习", "TD-MPC2", "几何泛化" ], "contribution": "InsertionWM以TD-MPC2学习潜在动力学、奖励、价值与策略先验，用MPPI在潜空间滚动规划。输入腕部96×96深度及24维本体/带噪位姿，输出六维末端增量；多装配共享模型，按各装配成功率调整初始状态课程。", "limitations": "全文只有仿真，实机部署仍列为未来难点，不能把仿真Panda当实体证据。留出对象同分布，不等于任意几何分布外泛化；AutoMate通用模型的原始划分不详，遮挡和小物体深度误差会影响比较与转移。", "codeStatus": "unknown", "codeStatusNote": "未核实此论文的完整官方实现；基础环境和 TD-MPC2 的公开代码不等于本论文开源。", "license": "未核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/html/2609.28258v1", "note": "§III 明确 experiments are conducted in simulation；§IV 说明 Panda/AutoMate 设置；摘要与 §VI 报告留出对象结果。" }, { "url": "https://arxiv.org/pdf/2609.28258", "note": "§VI-A：最后10个ID作同分布留出；训练90装配。" }, { "url": "https://arxiv.org/pdf/2609.28258", "note": "Fig.5 / §VI-B：79.78%对70.31%；通用零样本56%。" }, { "url": "https://arxiv.org/pdf/2609.28258", "note": "§VII：实机部署尚面临仿真差距、深度和实时性挑战。" } ], "tier": "recent", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对范围为公开原文、平台型号与所列来源；性能数字均为作者报告，未独立复现。", "year": 2026, "directions": [ "世界模型", "操作与抓取", "强化学习" ], "robotFilters": [], "original": { "id": "arxiv-2609.28258", "originalSourceUrl": "https://arxiv.org/pdf/2609.28258v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.28258v1", "finalDownloadUrl": "https://arxiv.org/pdf/2609.28258", "sha256": "574d3c826735bb20fac4fa70aee84e7a8248394f62e75411828b04ddcfcc47c1", "bytes": 3383401, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:33:21.870882+00:00", "publicationDate": "2026/09/23", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.28258", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.28258", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.28258", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "Generalizable Robotic Insertion with World Models", "sourceVersion": "2609.28258v1", "sourceVersionDate": "2026/09/23", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.613409+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.28258" }, "originalAnalysis": { "id": "arxiv-2609.28258", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090123+00:00", "sourceUrl": "https://arxiv.org/pdf/2609.28258", "sectionsRead": [ "§III–V", "§VI-A–B", "Figs.5–7", "§VII" ], "methodsZh": "InsertionWM以TD-MPC2学习潜在动力学、奖励、价值与策略先验，用MPPI在潜空间滚动规划。输入腕部96×96深度及24维本体/带噪位姿，输出六维末端增量；多装配共享模型，按各装配成功率调整初始状态课程。", "experimentsZh": "100套双零件装配资产、仿真Franka Panda；90训练、按ID留出10套同分布未见几何。专用策略各测5种子、每次100回合，通用模型每资产100次；与AutoMate、无深度、不同视角和课程等比较。", "resultsZh": "作者报告专用策略平均79.78%，AutoMate70.31%；90装配通用模型已见约55%、留出约56%。深度和课程明显帮助样本效率，专用模型训练200万步，AutoMate为5000万步。", "limitationsZh": "全文只有仿真，实机部署仍列为未来难点，不能把仿真Panda当实体证据。留出对象同分布，不等于任意几何分布外泛化；AutoMate通用模型的原始划分不详，遮挡和小物体深度误差会影响比较与转移。", "reproductionZh": "复现需AutoMate资产、奖励和预抓取流程，注意训练仍用CAD构建仿真以及带噪插座位姿/目标位置。模型500万参数，通用训练约三天RTX3090；无部署CAD不等于整个训练流程无CAD，本轮未运行。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "§VI-A", "note": "最后10个ID作同分布留出；训练90装配。" }, { "section/page": "Fig.5 / §VI-B", "note": "79.78%对70.31%；通用零样本56%。" }, { "section/page": "§VII", "note": "实机部署尚面临仿真差距、深度和实时性挑战。" } ], "corrections": { "experimentType": "sim", "robots": [], "robotNote": "Franka Panda只在本论文仿真环境中出现。" } }, "analysisVerifiedAt": "2026-10-04T13:50:35.090123+00:00" }, { "id": "arxiv-2609.28660", "title": "Morphometric Imitation: From Morphology and Contact Aware Hand Retargeting to Sim-to-Real Visuomotor Policy", "shortTitle": "Morphometric Imitation", "date": "2026-09-23", "updatedDate": "2026-09-25", "category": "灵巧操作 / 人类示范 / Sim-to-Real", "freshness": "最新预印本", "summary": "三阶段先用形态计量优化改变MANO掌指比例、恢复人手接触，再映射骨架并联合求臂手逆运动学；随后ManiSkill中的残差PPO修正参考关节轨迹，以物体位姿和接触奖励取得动力学可行性，最后把特权教师蒸馏成点云流匹…", "whyUseful": "复现需MANO/GRAB许可、手部URDF对应关系、接触容差、残差教师及512点点云预处理。策略不读RGB，所报180帧/秒重定向不含IK；应分别复现几何、物理和实机三阶段，而非直接回放人手关节。", "paperUrl": "https://arxiv.org/abs/2609.28660", "projectUrl": "https://morphometricimitation.github.io/", "codeUrl": "https://github.com/tsadja/morphometric", "trainingStatus": "占位仓库，待开源", "trainingNote": "官方 GitHub 目前仅 README/assets，明确写 Code will be released soon。", "caveats": "仓库存在不代表训练实现可用；89.3% 为指定对象与任务集的作者结果。", "url": "https://arxiv.org/abs/2609.28660", "status": "占位仓库，待开源", "titleZh": "形态度量模仿：从形态与接触感知的手部动作重定向，到仿真迁移现实的视觉运动策略", "abstractZh": "三阶段先用形态计量优化改变MANO掌指比例、恢复人手接触，再映射骨架并联合求臂手逆运动学；随后ManiSkill中的残差PPO修正参考关节轨迹，以物体位姿和接触奖励取得动力学可行性，最后把特权教师蒸馏成点云流匹配策略。桌面避碰贯穿重定向与训练。\n十段GRAB动捕轨迹分别测三指Dex3、四指Allegro、五指Sharpa，与五种基线比较；动态策略每类别2048仿真回合。实机仅KUKA iiwa14加Sharpa Wave，每类三物体，每物体九个固定网格初位加一个随机初位，共300次；要求抬升五厘米、完成轨迹并稳持五秒。", "experimentType": "both", "experimentNote": "十段GRAB动捕轨迹分别测三指Dex3、四指Allegro、五指Sharpa，与五种基线比较；动态策略每类别2048仿真回合。实机仅KUKA iiwa14加Sharpa Wave，每类三物体，每物体九个固定网格初位加一个随机初位，共300次；要求抬升五厘米、完成轨迹并稳持五秒。", "robots": [ "KUKA iiwa14 + Sharpa Wave", "Dex3（仿真）", "Allegro Hand（仿真）", "Sharpa（仿真）" ], "robotNote": "Dex3 和 Allegro 仅核实仿真实验；Sharpa Wave 为实机手，iiwa14 为搭载机械臂。", "tags": [ "灵巧操作", "人类示范", "动作重定向", "残差强化学习", "Sim-to-Real" ], "contribution": "三阶段先用形态计量优化改变MANO掌指比例、恢复人手接触，再映射骨架并联合求臂手逆运动学；随后ManiSkill中的残差PPO修正参考关节轨迹，以物体位姿和接触奖励取得动力学可行性，最后把特权教师蒸馏成点云流匹配策略。桌面避碰贯穿重定向与训练。", "limitations": "每物体类别独立训练策略，并非统一任意物体模型；数据只有十条动捕轨迹，未证实单目海量视频扩展。透明酒杯加入乒乓球才获得部分深度观测；定位区域和未覆盖几何形状仍导致失败，另外两种手没有实机测试。Sharpa两项接触F1基线差的置信区间包含零。", "codeStatus": "pending", "codeStatusNote": "官方仓库仅 README/assets，并明确 Code will be released soon。", "license": "未核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/html/2609.28660v2", "note": "§VII-A 明确 Dex3、Allegro、Sharpa 与 KUKA iiwa14；§VII-F 和 §VIII 给出实机范围与局限。" }, { "url": "https://github.com/tsadja/morphometric", "note": "官方仓库当前为占位内容，明确代码将发布。" }, { "url": "https://arxiv.org/pdf/2609.28660v2", "note": "VII-A; Appendix D：单一实机臂手系统，30物体300次，酒杯加乒乓球。" }, { "url": "https://arxiv.org/pdf/2609.28660v2", "note": "V; Tables III–IV：残差RL联合物体位姿和接触；固定配方对照。" }, { "url": "https://arxiv.org/pdf/2609.28660v2", "note": "IV-D; Figure 9：180FPS排除IK；学生用点云，不输入RGB。" }, { "url": "https://arxiv.org/pdf/2609.28660v2", "note": "v2 VII-A; Appendix D：已核对下载v2：每物体十次包括九个固定网格位置及一次随机，不是十个随机位置。" } ], "tier": "recent", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对范围为公开原文、平台型号与所列来源；性能数字均为作者报告，未独立复现。", "year": 2026, "directions": [ "操作与抓取", "灵巧手", "强化学习", "模仿学习" ], "robotFilters": [ "Sharpa Wave", "Allegro Hand", "Sharpa（型号未注明）" ], "original": { "id": "arxiv-2609.28660", "originalSourceUrl": "https://arxiv.org/pdf/2609.28660v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.28660v2", "finalDownloadUrl": "https://arxiv.org/pdf/2609.28660", "sha256": "abaee0b7796a4ab0fd4542b7f8fbbf2378fa1127358026596abfd4ca7d3900da", "bytes": 11523336, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:33:22.104666+00:00", "publicationDate": "2026/09/23", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.28660", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.28660", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.28660", "pages": 24, "sourceIdentityStatus": "title_match", "sourceTitle": "Morphometric Imitation: From Morphology and Contact Aware Hand Retargeting to Sim-to-Real Visuomotor Policy", "sourceVersion": "2609.28660v2", "sourceVersionDate": "2026/09/25", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.609202+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.28660" }, "originalAnalysis": { "id": "arxiv-2609.28660", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458238+00:00", "sourceUrl": "https://arxiv.org/pdf/2609.28660v2", "sectionsRead": [ "IV-A–D", "V Dynamic Retargeting", "VI Distillation", "VII-A–F", "Tables II–IV", "Appendix D", "Downloaded latest-version methods/results cross-check" ], "methodsZh": "三阶段先用形态计量优化改变MANO掌指比例、恢复人手接触，再映射骨架并联合求臂手逆运动学；随后ManiSkill中的残差PPO修正参考关节轨迹，以物体位姿和接触奖励取得动力学可行性，最后把特权教师蒸馏成点云流匹配策略。桌面避碰贯穿重定向与训练。", "experimentsZh": "十段GRAB动捕轨迹分别测三指Dex3、四指Allegro、五指Sharpa，与五种基线比较；动态策略每类别2048仿真回合。实机仅KUKA iiwa14加Sharpa Wave，每类三物体，每物体九个固定网格初位加一个随机初位，共300次；要求抬升五厘米、完成轨迹并稳持五秒。", "resultsZh": "作者报告三手接触F1均提升至少八点；相同残差学习下Sharpa成功率91.8%，最强基线56.5%。蒸馏学生仿真93.8%，真实268/300即89.3%，未观察硬桌碰撞失败；这些是给定试验结果，不是安全保证。", "limitationsZh": "每物体类别独立训练策略，并非统一任意物体模型；数据只有十条动捕轨迹，未证实单目海量视频扩展。透明酒杯加入乒乓球才获得部分深度观测；定位区域和未覆盖几何形状仍导致失败，另外两种手没有实机测试。Sharpa两项接触F1基线差的置信区间包含零。", "reproductionZh": "复现需MANO/GRAB许可、手部URDF对应关系、接触容差、残差教师及512点点云预处理。策略不读RGB，所报180帧/秒重定向不含IK；应分别复现几何、物理和实机三阶段，而非直接回放人手关节。", "experimentType": "both", "robots": [ "KUKA iiwa14 + Sharpa Wave", "Dex3（仿真）", "Allegro Hand（仿真）", "Sharpa（仿真）" ], "evidenceNotes": [ { "section": "VII-A; Appendix D", "note": "单一实机臂手系统，30物体300次，酒杯加乒乓球。" }, { "section": "V; Tables III–IV", "note": "残差RL联合物体位姿和接触；固定配方对照。" }, { "section": "IV-D; Figure 9", "note": "180FPS排除IK；学生用点云，不输入RGB。" }, { "section": "v2 VII-A; Appendix D", "note": "已核对下载v2：每物体十次包括九个固定网格位置及一次随机，不是十个随机位置。" } ], "corrections": [ "明确实机KUKA iiwa14 + Sharpa Wave；Dex3/Allegro仅仿真。" ], "sourceVersion": "2609.28660v2", "originalSha256": "abaee0b7796a4ab0fd4542b7f8fbbf2378fa1127358026596abfd4ca7d3900da", "additionalSourcesRead": [ "https://arxiv.org/html/2609.28660v1" ] }, "analysisVerifiedAt": "2026-10-04T13:45:34.458238+00:00" }, { "id": "arxiv-2609.23968", "title": "Opt2VLA: Force-Aware Vision-Language-Action for Contact-Rich Humanoid Whole-Body Manipulation", "shortTitle": "Opt2VLA", "date": "2026-09-21", "category": "VLA / 人形全身操作 / 力控制", "freshness": "最新预印本", "summary": "Opt2VLA把接触力目标显式加入GR00T-N1.7输出，与手脚运动目标一起交给任务专用全身RL控制器。轨迹优化生成满足动力学的动作、接触力及关节力矩参考，力矩参考仅训练时提供；实测力矩历史经独立编码器输入V…", "whyUseful": "需分别保存轨迹优化约束、控制器奖励、关节实测力矩标定和提示力区间；复现独立力矩token与简单拼接的消融。评估同时统计任务门槛、指令力匹配、真实力误差，防止只报告几何完成。", "paperUrl": "https://arxiv.org/abs/2609.23968", "projectUrl": "https://opt2vla.github.io/", "codeUrl": null, "trainingStatus": "未核实公开训练代码", "trainingNote": "项目页有 Code 文本链接，但本次点击仍回到项目页，未核实独立公开代码仓库。", "caveats": "底层控制器仍为任务专用，不应称为已解决通用人形控制；实机执行需要相应硬件和安全流程。", "url": "https://arxiv.org/abs/2609.23968", "status": "未核实公开训练代码", "titleZh": "Opt2VLA：面向接触密集型人形机器人全身操作的力感知视觉—语言—动作框架", "abstractZh": "Opt2VLA把接触力目标显式加入GR00T-N1.7输出，与手脚运动目标一起交给任务专用全身RL控制器。轨迹优化生成满足动力学的动作、接触力及关节力矩参考，力矩参考仅训练时提供；实测力矩历史经独立编码器输入VLA，另预测未来力矩作辅助监督。\nAgility Robotics Digit，48千克、30自由度20主动关节，做擦桌、推盒入架、双手提盒。九个VLA配置每个90个仿真回合；硬件低层每任务30次，端到端按三任务×三力提示×五次共45次。", "experimentType": "both", "experimentNote": "Agility Robotics Digit，48千克、30自由度20主动关节，做擦桌、推盒入架、双手提盒。九个VLA配置每个90个仿真回合；硬件低层每任务30次，端到端按三任务×三力提示×五次共45次。", "robots": [ "Agility Robotics Digit" ], "robotNote": "论文 §IV-A 明确平台为 Digit，不能替换成 Unitree G1。", "tags": [ "VLA", "人形机器人", "全身操作", "力控制", "轨迹优化" ], "contribution": "Opt2VLA把接触力目标显式加入GR00T-N1.7输出，与手脚运动目标一起交给任务专用全身RL控制器。轨迹优化生成满足动力学的动作、接触力及关节力矩参考，力矩参考仅训练时提供；实测力矩历史经独立编码器输入VLA，另预测未来力矩作辅助监督。", "limitations": "82.2%不是45次实机成功率；实机擦桌虽相关性高仍有7N偏差。提示对应固定训练力区间且场景外观来自训练家族，泛化有限；三个底层控制器各自专用，尚非任意接触任务通用控制。", "codeStatus": "unknown", "codeStatusNote": "项目页有 Code 链接但未核实独立实现仓库；不能据按钮存在判定已开源。", "license": "未核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/html/2609.23968v1", "note": "§IV-A 明确仿真和实机 Digit；§IV-D 分开报告控制器与完整系统的硬件实验。" }, { "url": "https://opt2vla.github.io/", "note": "官方项目展示运动/力接口与硬件结果；Code 入口未提供已核实的实现。" }, { "url": "https://arxiv.org/pdf/2609.23968", "note": "III; IV-A：Digit硬件、显式力接口及特权力矩。" }, { "url": "https://arxiv.org/pdf/2609.23968", "note": "Table III; IV-D：90仿真与45实机分母、联合指标和力MAE。" }, { "url": "https://arxiv.org/pdf/2609.23968v1", "note": "III; IV-A：Digit硬件、显式力接口及特权力矩。" }, { "url": "https://arxiv.org/pdf/2609.23968v1", "note": "Table III; IV-D：90仿真与45实机分母、联合指标和力MAE。" } ], "tier": "recent", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对范围为公开原文、平台型号与所列来源；性能数字均为作者报告，未独立复现。", "year": 2026, "directions": [ "视觉语言动作", "运动控制", "操作与抓取", "人形机器人" ], "robotFilters": [ "Agility Robotics Digit" ], "original": { "id": "arxiv-2609.23968", "originalSourceUrl": "https://arxiv.org/pdf/2609.23968v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.23968v1", "finalDownloadUrl": "https://arxiv.org/pdf/2609.23968", "sha256": "a4c1f70308b89a38ff4967b5b57b92a5a3cc59b97f76236ed1421b20746ea124", "bytes": 17021553, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:33:36.971210+00:00", "publicationDate": "2026/09/21", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.23968", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.23968", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.23968", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "Opt2VLA: Force-Aware Vision-Language-Action for Contact-Rich Humanoid Whole-Body Manipulation", "sourceVersion": "2609.23968v1", "sourceVersionDate": "2026/09/21", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.631188+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.23968" }, "originalAnalysis": { "id": "arxiv-2609.23968", "sourceUrl": "https://arxiv.org/pdf/2609.23968v1", "sectionsRead": [ "III-A–C", "IV-A–D", "V", "Tables I–III; Figs 6–7" ], "methodsZh": "Opt2VLA把接触力目标显式加入GR00T-N1.7输出，与手脚运动目标一起交给任务专用全身RL控制器。轨迹优化生成满足动力学的动作、接触力及关节力矩参考，力矩参考仅训练时提供；实测力矩历史经独立编码器输入VLA，另预测未来力矩作辅助监督。", "experimentsZh": "Agility Robotics Digit，48千克、30自由度20主动关节，做擦桌、推盒入架、双手提盒。九个VLA配置每个90个仿真回合；硬件低层每任务30次，端到端按三任务×三力提示×五次共45次。", "resultsZh": "默认配置仿真联合成功82.2%，运动状态基线40.0%；联合标准同时要求任务完成、语言力区间及命令—实测均值误差。实机低层力MAE擦桌7.0N、推盒3.9N、提盒1.7N，端到端显示轻/中/强力指令可调制实际作用力。", "limitationsZh": "82.2%不是45次实机成功率；实机擦桌虽相关性高仍有7N偏差。提示对应固定训练力区间且场景外观来自训练家族，泛化有限；三个底层控制器各自专用，尚非任意接触任务通用控制。", "reproductionZh": "需分别保存轨迹优化约束、控制器奖励、关节实测力矩标定和提示力区间；复现独立力矩token与简单拼接的消融。评估同时统计任务门槛、指令力匹配、真实力误差，防止只报告几何完成。", "experimentType": "both", "robots": [ "Agility Robotics Digit" ], "corrections": [ "82.2%属于仿真VLA联合指标；实机结果应引用45次力调制与独立低层力误差。" ], "evidenceNotes": [ { "note": "Digit硬件、显式力接口及特权力矩。", "section": "III; IV-A" }, { "note": "90仿真与45实机分母、联合指标和力MAE。", "section": "Table III; IV-D" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2609.23968v1", "originalSha256": "a4c1f70308b89a38ff4967b5b57b92a5a3cc59b97f76236ed1421b20746ea124" }, "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2609.24682", "title": "Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies", "shortTitle": "THAW-VLA", "date": "2026-09-21", "updatedDate": "2026-09-22", "category": "VLA / 世界模型蒸馏", "freshness": "最新预印本", "summary": "把冻结世界模型的每相机图像令牌均值离线缓存，再用小投影头将0.8B学生VLA的视觉特征与教师做余弦对齐；与流匹配动作损失共同训练。部署移除教师、缓存和投影器，保持原学生一次主干前向与四步流解码。", "whyUseful": "保留教师特征提取层、逐视角缓存键及0.5对齐权重；固定数据、学生和推理流程做成对消融，分别报告训练与评估种子。代码与训练效果本次未执行验证。", "paperUrl": "https://arxiv.org/abs/2609.24682", "projectUrl": "https://thaw-vla.trung-dt.com/", "codeUrl": "https://github.com/trungdt880/THAW-VLA", "modelUrl": "https://huggingface.co/collections/termanteus/thaw-vla", "trainingStatus": "训练代码已核实", "trainingNote": "README 默认配置使用 4 GPU，教师 Cosmos3-Nano 约 33 GB；HF 两个 0.8B 模型文件页可见 final_model.pt、config 和归一化统计，各约 2.54 GB。未实际下载或运行。", "caveats": "GitHub README 仍称权重 private/request access，但当前 HF 文件列表可公开读取；下载权限未实测。论文的低推理显存不是完整训练显存。", "extraSources": [ "https://huggingface.co/termanteus/THAW-VLA-Qwen3.5-0.8B-LIBERO/tree/main", "https://huggingface.co/termanteus/THAW-VLA-Qwen3.5-0.8B-Robocasa-GR1/tree/main" ], "url": "https://arxiv.org/abs/2609.24682", "status": "训练代码已核实", "titleZh": "像世界模型一样思考，像 VLA 一样行动：将世界模型表征蒸馏到紧凑机器人策略中", "abstractZh": "把冻结世界模型的每相机图像令牌均值离线缓存，再用小投影头将0.8B学生VLA的视觉特征与教师做余弦对齐；与流匹配动作损失共同训练。部署移除教师、缓存和投影器，保持原学生一次主干前向与四步流解码。\nLIBERO每任务50回合，RoboCasa-GR1的24环境每个20回合；本方法四次评估由两种子×两GPU组成。AgileX Nero单臂水果/鸡蛋搬运及TRIP-Bag双臂交接，每策略每条件30次；比较无蒸馏学生、4B策略及不同教师、层级和规模。", "experimentType": "both", "experimentNote": "LIBERO每任务50回合，RoboCasa-GR1的24环境每个20回合；本方法四次评估由两种子×两GPU组成。AgileX Nero单臂水果/鸡蛋搬运及TRIP-Bag双臂交接，每策略每条件30次；比较无蒸馏学生、4B策略及不同教师、层级和规模。", "robots": [ "AgileX Nero", "TRIP-Bag", "Fourier GR1（仿真）" ], "robotNote": "AgileX Nero 和 TRIP-Bag 为实机；GR1 是 RoboCasa-GR1 仿真具身。", "tags": [ "VLA", "世界模型", "知识蒸馏", "轻量策略", "双臂操作" ], "contribution": "把冻结世界模型的每相机图像令牌均值离线缓存，再用小投影头将0.8B学生VLA的视觉特征与教师做余弦对齐；与流匹配动作损失共同训练。部署移除教师、缓存和投影器，保持原学生一次主干前向与四步流解码。", "limitations": "提升幅度有限，表内多种外部模型成绩引用其他论文，不能视为全部同条件重跑。GPU与采样影响成绩；教师表征对齐不证明学生学得准确动力学。真实仅两平台三任务，易滑物体与交接仍困难。", "codeStatus": "open", "codeStatusNote": "官方仓库包含特征预计算、训练、评估实现，许可证可读；未实际运行训练。", "license": "MIT（保留 StarVLA 归属和仓库附加说明）", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/html/2609.24682v2", "note": "摘要、§IV-A/IV-D 明确两类仿真与 AgileX Nero、TRIP-Bag 实机；GR1 型号在基准说明中。" }, { "url": "https://github.com/trungdt880/THAW-VLA", "note": "README 和目录提供预计算、训练及 LIBERO/RoboCasa-GR1 评估入口。" }, { "url": "https://github.com/trungdt880/THAW-VLA/blob/main/LICENSE", "note": "可读 MIT 许可，含 StarVLA 归属及保留上游提交等附加说明。" }, { "url": "https://arxiv.org/pdf/2609.24682", "note": "§III-B–C：离线缓存、Cosmos3-Nano第24层特征及部署移除模块。" }, { "url": "https://arxiv.org/pdf/2609.24682", "note": "Tables I–III; §IV-A：四评估运行、30次实机试验、三任务成功率。" } ], "tier": "recent", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对范围为公开原文、平台型号与所列来源；性能数字均为作者报告，未独立复现。", "year": 2026, "directions": [ "视觉语言动作", "世界模型", "操作与抓取", "模仿学习" ], "robotFilters": [ "AgileX Nero", "TRIP-Bag", "Fourier GR-1" ], "original": { "id": "arxiv-2609.24682", "originalSourceUrl": "https://arxiv.org/pdf/2609.24682v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.24682v2", "finalDownloadUrl": "https://arxiv.org/pdf/2609.24682", "sha256": "3abb895e9911cf97a2a279ec51d0c8282295d5ad7e2e4aa4f7bf0c10289a9c4e", "bytes": 5372281, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:33:22.025430+00:00", "publicationDate": "2026/09/21", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.24682", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.24682", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.24682", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies", "sourceVersion": "2609.24682v2", "sourceVersionDate": "2026/09/22", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.618573+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.24682" }, "originalAnalysis": { "id": "arxiv-2609.24682", "sourceUrl": "https://arxiv.org/pdf/2609.24682", "sectionsRead": [ "§III", "§IV-A–E", "Tables I–VI", "§V" ], "methodsZh": "把冻结世界模型的每相机图像令牌均值离线缓存，再用小投影头将0.8B学生VLA的视觉特征与教师做余弦对齐；与流匹配动作损失共同训练。部署移除教师、缓存和投影器，保持原学生一次主干前向与四步流解码。", "experimentsZh": "LIBERO每任务50回合，RoboCasa-GR1的24环境每个20回合；本方法四次评估由两种子×两GPU组成。AgileX Nero单臂水果/鸡蛋搬运及TRIP-Bag双臂交接，每策略每条件30次；比较无蒸馏学生、4B策略及不同教师、层级和规模。", "resultsZh": "作者报告LIBERO均值95.3%升至97.9%，GR1由48.2%升至50.5%。实机水果、鸡蛋、双臂交接由83.3/46.7/40.0%升至93.3/60.0/46.7%，但后两项仍低于4B参照；每次需所有物体入容器才算成功。", "limitationsZh": "提升幅度有限，表内多种外部模型成绩引用其他论文，不能视为全部同条件重跑。GPU与采样影响成绩；教师表征对齐不证明学生学得准确动力学。真实仅两平台三任务，易滑物体与交接仍困难。", "reproductionZh": "保留教师特征提取层、逐视角缓存键及0.5对齐权重；固定数据、学生和推理流程做成对消融，分别报告训练与评估种子。代码与训练效果本次未执行验证。", "experimentType": "both", "robots": [ "AgileX Nero", "TRIP-Bag", "Fourier GR1（仿真）" ], "evidenceNotes": [ { "section/page": "§III-B–C", "note": "离线缓存、Cosmos3-Nano第24层特征及部署移除模块。" }, { "section/page": "Tables I–III; §IV-A", "note": "四评估运行、30次实机试验、三任务成功率。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2609.22895", "title": "H-VLA: Hierarchical Vision-Language-Action Model with Key-Action Reasoning and Motion Planning in a Unified Action Space", "shortTitle": "H-VLA", "date": "2026-09-19", "category": "VLA / 层级规划 / 跨具身", "freshness": "最新预印本", "summary": "H-VLA用Prismatic-7B视觉语言骨干提取认知特征，先由扩散模块生成末端位姿与夹爪组成的关键动作，再由另一扩散模块生成密集动作段。状态、子目标及动作统一到已标定第三人称相机坐标；关键帧按夹爪变化及轨迹…", "whyUseful": "复现需保留跨数据集标定、单/双臂掩码和关键动作标签；八H100预训练约三天，真机适配四H100约一天。附录列逐任务试次数、无提前成功终止的仿真协议及PiPER硬件，4090端到端推理约103毫秒。", "paperUrl": "https://arxiv.org/abs/2609.22895", "projectUrl": null, "codeUrl": null, "trainingStatus": "未核实公开训练代码", "trainingNote": "已查论文正文/附录和作者相关结果，未发现可核实的官方实现链接。", "caveats": "不同基线的实机试验次数、覆盖的 OOD 设置不同；不能把单一平均值当完全等价对照。", "extraSources": [ "https://arxiv.org/html/2609.22895v1" ], "url": "https://arxiv.org/abs/2609.22895", "status": "未核实公开训练代码", "titleZh": "H-VLA：在统一动作空间中结合关键动作推理与运动规划的层级视觉—语言—动作模型", "abstractZh": "H-VLA用Prismatic-7B视觉语言骨干提取认知特征，先由扩散模块生成末端位姿与夹爪组成的关键动作，再由另一扩散模块生成密集动作段。状态、子目标及动作统一到已标定第三人称相机坐标；关键帧按夹爪变化及轨迹终点自动标注，先重关键动作预训练，再等权联合微调。\n混合约18.3万轨迹预训练，分别微调后测Google/WidowX的SimplerEnv。真实双臂Agilex PiPER用三任务共150示范；H-VLA及π系列各228次，分ID、位置OOD和场景/物体OOD，CogACT与MolmoAct2只做较少的ID/位置试验。", "experimentType": "both", "experimentNote": "混合约18.3万轨迹预训练，分别微调后测Google/WidowX的SimplerEnv。真实双臂Agilex PiPER用三任务共150示范；H-VLA及π系列各228次，分ID、位置OOD和场景/物体OOD，CogACT与MolmoAct2只做较少的ID/位置试验。", "robots": [ "Agilex PiPER（双臂Cobot Magic式平台）", "Google Robot（SimplerEnv仿真）", "WidowX（SimplerEnv仿真）" ], "robotNote": "实机为两台 PiPER 从臂，Cobot Magic 风格配置；Google Robot/WidowX 是仿真基准原文名称，未推断更细子型号。", "tags": [ "VLA", "层级规划", "关键动作", "跨具身", "空间泛化", "双臂操作" ], "contribution": "H-VLA用Prismatic-7B视觉语言骨干提取认知特征，先由扩散模块生成末端位姿与夹爪组成的关键动作，再由另一扩散模块生成密集动作段。状态、子目标及动作统一到已标定第三人称相机坐标；关键帧按夹爪变化及轨迹终点自动标注，先重关键动作预训练，再等权联合微调。", "limitations": "基线预训练数据、相机输入及评测数量不同，CogACT仅一个视角。关键标签会漏掉中间接触事件；相机外参误差敏感性未系统测，灵巧接触与变形操作仍困难。训练数据中的Franka不能算新增实机验证。", "codeStatus": "unknown", "codeStatusNote": "未核实论文对应官方公开实现；未发现仓库不等于明确不开放代码。", "license": "未核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/html/2609.22895v1", "note": "摘要、§4 与附录 C/D：SimplerEnv、Agilex PiPER、Cobot Magic 风格实机及比较协议。" }, { "url": "https://arxiv.org/html/2609.22895v1", "note": "3; Appendix A：统一相机坐标依赖标定；单臂七维、模型统一十四维。" }, { "url": "https://arxiv.org/html/2609.22895v1", "note": "Appendix D.1; D.3; Tables 11–12：PiPER真机；228与80试次配置不同，宏平均。" }, { "url": "https://arxiv.org/html/2609.22895v1", "note": "4.3; Table 3：7.7B总参数，4090 103ms含腕相机编码。" } ], "tier": "recent", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对范围为公开原文、平台型号与所列来源；性能数字均为作者报告，未独立复现。", "year": 2026, "directions": [ "视觉语言动作", "操作与抓取" ], "robotFilters": [ "AgileX PiPER", "Google Robot", "WidowX" ], "original": { "id": "arxiv-2609.22895", "originalSourceUrl": "https://arxiv.org/pdf/2609.22895v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.22895v1", "finalDownloadUrl": "https://arxiv.org/pdf/2609.22895", "sha256": "41125aece3618a5f2b9a7843acce5fb33d55450ab2057f482d75fa36551dec9a", "bytes": 22594798, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:33:36.891794+00:00", "publicationDate": "2026/09/19", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.22895", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.22895", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.22895", "pages": 27, "sourceIdentityStatus": "title_match", "sourceTitle": "H-VLA: Hierarchical Vision-Language-Action Model with Key-Action Reasoning and Motion Planning in a Unified Action Space", "sourceVersion": "2609.22895v1", "sourceVersionDate": "2026/09/19", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.654249+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.22895" }, "originalAnalysis": { "id": "arxiv-2609.22895", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458241+00:00", "sourceUrl": "https://arxiv.org/html/2609.22895v1", "sectionsRead": [ "3.1–3.6", "4.1–4.4", "5 Limitations", "Appendix A", "Appendix C.2–C.3", "Appendix D.1–D.3" ], "methodsZh": "H-VLA用Prismatic-7B视觉语言骨干提取认知特征，先由扩散模块生成末端位姿与夹爪组成的关键动作，再由另一扩散模块生成密集动作段。状态、子目标及动作统一到已标定第三人称相机坐标；关键帧按夹爪变化及轨迹终点自动标注，先重关键动作预训练，再等权联合微调。", "experimentsZh": "混合约18.3万轨迹预训练，分别微调后测Google/WidowX的SimplerEnv。真实双臂Agilex PiPER用三任务共150示范；H-VLA及π系列各228次，分ID、位置OOD和场景/物体OOD，CogACT与MolmoAct2只做较少的ID/位置试验。", "resultsZh": "作者报告仿真三个分割均值91%、84%、81%；真机三类条件的任务宏平均93%、83%、66%。位置OOD比最强所测对照高47点，但场景变化中的LiftPot仍落后π系列，且这些均值不是按所有试次合并的比例。", "limitationsZh": "基线预训练数据、相机输入及评测数量不同，CogACT仅一个视角。关键标签会漏掉中间接触事件；相机外参误差敏感性未系统测，灵巧接触与变形操作仍困难。训练数据中的Franka不能算新增实机验证。", "reproductionZh": "复现需保留跨数据集标定、单/双臂掩码和关键动作标签；八H100预训练约三天，真机适配四H100约一天。附录列逐任务试次数、无提前成功终止的仿真协议及PiPER硬件，4090端到端推理约103毫秒。", "experimentType": "both", "robots": [ "Agilex PiPER（双臂Cobot Magic式平台）", "Google Robot（SimplerEnv仿真）", "WidowX（SimplerEnv仿真）" ], "evidenceNotes": [ { "section": "3; Appendix A", "note": "统一相机坐标依赖标定；单臂七维、模型统一十四维。" }, { "section": "Appendix D.1; D.3; Tables 11–12", "note": "PiPER真机；228与80试次配置不同，宏平均。" }, { "section": "4.3; Table 3", "note": "7.7B总参数，4090 103ms含腕相机编码。" } ], "corrections": [ "Agilex实机具体为PiPER；不要把预训练数据平台全部标作实机试验。" ], "sourceVersion": "2609.22895v1", "originalSha256": "41125aece3618a5f2b9a7843acce5fb33d55450ab2057f482d75fa36551dec9a" }, "analysisVerifiedAt": "2026-10-04T13:45:34.458241+00:00" }, { "id": "arxiv-2609.18293", "title": "Function-Preserving Data Generation for Zero-Shot Real-to-Sim-to-Real Manipulation", "shortTitle": "Function-Preserving R2S2R", "date": "2026-09-16", "category": "Sim-to-Real / 合成数据", "freshness": "最新预印本", "summary": "先由现实影像重建工作区与物体，再标注任务关键界面，在保留接触面和孔/头部约束下拉伸、弯曲非功能几何。顶点对应关系同步传递任务位姿和碰撞代理，用具特权状态的脚本专家批量生成示范，视觉策略只读取图像与本体状态。", "whyUseful": "复现应同时保存变形约束、任务坐标、碰撞代理和随机化种子，区分脚本专家特权状态与视觉策略输入。表I涵盖标定、光照等随机化；需要公开分任务实机试验数和失败明细，本轮未据汇总均值补造这些数。", "paperUrl": "https://arxiv.org/abs/2609.18293", "projectUrl": "https://fpsa-r2s2r.github.io/", "codeUrl": "https://github.com/FPSA-r2s2r/FPSA-r2s2r", "trainingStatus": "占位仓库，待开源", "trainingNote": "官方仓库仅 README；明确代码、资产和文档正在准备公开。", "caveats": "项目页仍使用 Anonymous Authors；作者身份以 arXiv 为准，不能据 Code 按钮认定可直接复现。", "url": "https://arxiv.org/abs/2609.18293", "status": "占位仓库，待开源", "titleZh": "面向零样本“现实—仿真—现实”机器人操作的功能保持式数据生成", "abstractZh": "先由现实影像重建工作区与物体，再标注任务关键界面，在保留接触面和孔/头部约束下拉伸、弯曲非功能几何。顶点对应关系同步传递任务位姿和碰撞代理，用具特权状态的脚本专家批量生成示范，视觉策略只读取图像与本体状态。\n仿真PyBullet生成每任务3000回合，训练任务专用策略后直接部署Franka Panda。五任务包括抓取、扳手拧紧、装配、单齿轮和三齿轮顺序取出；实机用未见3D打印形状及静/动态干扰，仿真每任务50留出环境比较缩放和资产检索。", "experimentType": "both", "experimentNote": "仿真PyBullet生成每任务3000回合，训练任务专用策略后直接部署Franka Panda。五任务包括抓取、扳手拧紧、装配、单齿轮和三齿轮顺序取出；实机用未见3D打印形状及静/动态干扰，仿真每任务50留出环境比较缩放和资产检索。", "robots": [ "Franka Panda" ], "robotNote": "实机配桌面 RealSense L515 与腕部鱼眼 RGB 相机。", "tags": [ "Real-to-Sim-to-Real", "合成数据", "几何增广", "接触操作", "装配" ], "contribution": "先由现实影像重建工作区与物体，再标注任务关键界面，在保留接触面和孔/头部约束下拉伸、弯曲非功能几何。顶点对应关系同步传递任务位姿和碰撞代理，用具特权状态的脚本专家批量生成示范，视觉策略只读取图像与本体状态。", "limitations": "未见物体仍保留人为定义的兼容功能界面，不能外推到界面改变或任意物体。无需遥操作轨迹不等于无人工作，仍需重建、功能标注与脚本；仿真资产基线中的不兼容结构亦可能影响公平性。", "codeStatus": "pending", "codeStatusNote": "官方仓库明确代码、资产和文档正在准备公开。", "license": "未核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/html/2609.18293v1", "note": "§IV-D 与 §V 明确 PyBullet 仿真、Franka Panda 实机、脚本专家和测试范围。" }, { "url": "https://github.com/FPSA-r2s2r/FPSA-r2s2r", "note": "官方仓库 Coming Soon 段落确认待发布。" }, { "url": "https://arxiv.org/pdf/2609.18293", "note": "§V-A：五任务、每任务3000合成训练回合，实机Panda零微调。" }, { "url": "https://arxiv.org/pdf/2609.18293", "note": "Fig.5 / §V-A：82%标准；74%动态干扰；68%动态光照。" }, { "url": "https://arxiv.org/pdf/2609.18293", "note": "§V-B：统一随机化与50仿真留出环境；关键接口兼容是任务前提。" } ], "tier": "recent", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对范围为公开原文、平台型号与所列来源；性能数字均为作者报告，未独立复现。", "year": 2026, "directions": [ "操作与抓取" ], "robotFilters": [ "Franka Panda" ], "original": { "id": "arxiv-2609.18293", "originalSourceUrl": "https://arxiv.org/pdf/2609.18293v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.18293v1", "finalDownloadUrl": "https://arxiv.org/pdf/2609.18293", "sha256": "18bb3e61c5fc614c23508a34edbf63b4570153f8d341d67dc5e8d4e280b6b2aa", "bytes": 2636149, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:33:36.504993+00:00", "publicationDate": "2026/09/16", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.18293", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.18293", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.18293", "pages": 9, "sourceIdentityStatus": "title_match", "sourceTitle": "Function-Preserving Data Generation for Zero-Shot Real-to-Sim-to-Real Manipulation", "sourceVersion": "2609.18293v1", "sourceVersionDate": "2026/09/16", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.658006+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.18293" }, "originalAnalysis": { "id": "arxiv-2609.18293", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090125+00:00", "sourceUrl": "https://arxiv.org/pdf/2609.18293", "sectionsRead": [ "§III–IV", "§V-A–C", "Figs.2,4–8", "Table I", "§VI" ], "methodsZh": "先由现实影像重建工作区与物体，再标注任务关键界面，在保留接触面和孔/头部约束下拉伸、弯曲非功能几何。顶点对应关系同步传递任务位姿和碰撞代理，用具特权状态的脚本专家批量生成示范，视觉策略只读取图像与本体状态。", "experimentsZh": "仿真PyBullet生成每任务3000回合，训练任务专用策略后直接部署Franka Panda。五任务包括抓取、扳手拧紧、装配、单齿轮和三齿轮顺序取出；实机用未见3D打印形状及静/动态干扰，仿真每任务50留出环境比较缩放和资产检索。", "resultsZh": "作者报告实机标准条件平均82%，动态物体干扰74%、动态照明68%。仿真接触密集任务本方法超过80%，两基线不高于65%；固定回合总量增加几何种类仍有收益。", "limitationsZh": "未见物体仍保留人为定义的兼容功能界面，不能外推到界面改变或任意物体。无需遥操作轨迹不等于无人工作，仍需重建、功能标注与脚本；仿真资产基线中的不兼容结构亦可能影响公平性。", "reproductionZh": "复现应同时保存变形约束、任务坐标、碰撞代理和随机化种子，区分脚本专家特权状态与视觉策略输入。表I涵盖标定、光照等随机化；需要公开分任务实机试验数和失败明细，本轮未据汇总均值补造这些数。", "experimentType": "both", "robots": [ "Franka Panda" ], "evidenceNotes": [ { "section/page": "§V-A", "note": "五任务、每任务3000合成训练回合，实机Panda零微调。" }, { "section/page": "Fig.5 / §V-A", "note": "82%标准；74%动态干扰；68%动态光照。" }, { "section/page": "§V-B", "note": "统一随机化与50仿真留出环境；关键接口兼容是任务前提。" } ] }, "analysisVerifiedAt": "2026-10-04T13:50:35.090125+00:00" }, { "id": "arxiv-2609.13053", "title": "Dynin-Robotics: Omnimodal Unified Diffusion Vision-Language-Action Model", "shortTitle": "Dynin-Robotics", "date": "2026-09-11", "category": "VLA / 统一世界模型", "freshness": "近期预印本", "summary": "在共享掩码扩散主干中序列化文字、视觉和量化动作，联合训练策略、下一图像、任务理解与目标图像预测。动作按块并行迭代解码；默认先生成目标，再条件化动作，也研究联合动作—视觉去噪和固定参考的候选重排。", "whyUseful": "区分已发布220k阶段检查点与后续任务模型，记录量化区间、目标混合比例、去噪步数和完整延迟；分别检验任务成功与视觉变化预测。未运行代码或独立复核作者成绩。", "paperUrl": "https://arxiv.org/abs/2609.13053", "projectUrl": "https://dynin.ai/robotics/", "codeUrl": "https://github.com/AIDASLab/Dynin-Robotics", "trainingStatus": "占位仓库，待开源", "trainingNote": "官方项目页写 Model and Code will be released soon；GitHub 目前主要是 README/assets。", "caveats": "项目页展示更早发布日期和 NeurIPS 2026 标签，本清单统一采用可核实 arXiv 初次提交日期。速度提升只针对论文指定的模型侧解码基线。", "url": "https://arxiv.org/abs/2609.13053", "status": "占位仓库，待开源", "titleZh": "Dynin-Robotics：全模态统一扩散视觉—语言—动作模型", "abstractZh": "在共享掩码扩散主干中序列化文字、视觉和量化动作，联合训练策略、下一图像、任务理解与目标图像预测。动作按块并行迭代解码；默认先生成目标，再条件化动作，也研究联合动作—视觉去噪和固定参考的候选重排。\n经OXE机器人继续预训练后，在LIBERO、七类LIBERO-Plus扰动和VLABench两任务评测；Franka Research 3测试水果搬运、分类、自由堆叠及指定颜色顺序堆叠。消融分别固定解码器改变训练目标、固定检查点改变推理组合。", "experimentType": "both", "experimentNote": "经OXE机器人继续预训练后，在LIBERO、七类LIBERO-Plus扰动和VLABench两任务评测；Franka Research 3测试水果搬运、分类、自由堆叠及指定颜色顺序堆叠。消融分别固定解码器改变训练目标、固定检查点改变推理组合。", "robots": [ "Franka Research 3" ], "robotNote": "实机型号由 §5.1 和 §5.2.4 明确；未把 OXE 预训练数据中的全部机器人当成本论文实机平台。", "tags": [ "VLA", "世界模型", "扩散模型", "多模态统一", "目标条件控制" ], "contribution": "在共享掩码扩散主干中序列化文字、视觉和量化动作，联合训练策略、下一图像、任务理解与目标图像预测。动作按块并行迭代解码；默认先生成目标，再条件化动作，也研究联合动作—视觉去噪和固定参考的候选重排。", "limitations": "已预留语音/力触觉接口但报告的机器人训练和策略解码未启用，不能称多传感器验证。真实仅单平台单工作区；近未来图像全图指标上复制原帧仍更强。任务理解只定性评估，视觉预测收益取决于推理组合。", "codeStatus": "pending", "codeStatusNote": "官方仓库明确 Code and model will be released soon。", "license": "未核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/html/2609.13053v1", "note": "摘要、§5.1、§5.2.4 和表 5 给出训练规模、仿真/实机、FR3 型号及任务边界。" }, { "url": "https://github.com/AIDASLab/Dynin-Robotics", "note": "仓库 Notifications 明确代码和模型待发布。" }, { "url": "https://arxiv.org/pdf/2609.13053", "note": "§3.2.2; §3.4：传感器接口未启用；目标条件和联合去噪不同推理路径。" }, { "url": "https://arxiv.org/pdf/2609.13053", "note": "Tables 2–3,5,9–11; §6：仿真/实机成绩、模型侧速度及视觉预测限制。" } ], "tier": "recent", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对范围为公开原文、平台型号与所列来源；性能数字均为作者报告，未独立复现。", "year": 2026, "directions": [ "视觉语言动作", "世界模型" ], "robotFilters": [ "Franka Research 3" ], "original": { "id": "arxiv-2609.13053", "originalSourceUrl": "https://arxiv.org/pdf/2609.13053v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2609.13053v1", "finalDownloadUrl": "https://arxiv.org/pdf/2609.13053", "sha256": "18662cabd0593a95cde4b859c0a62334d2e139fc0d88895125abc4f003ca1ac5", "bytes": 40846917, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:33:48.105129+00:00", "publicationDate": "2026/09/11", "publicationDateSourceUrl": "https://arxiv.org/abs/2609.13053", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2609.13053", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2609.13053", "pages": 36, "sourceIdentityStatus": "title_match", "sourceTitle": "Dynin-Robotics: Omnimodal Unified Diffusion Vision-Language-Action Model", "sourceVersion": "2609.13053v1", "sourceVersionDate": "2026/09/11", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.696411+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2609.13053" }, "originalAnalysis": { "id": "arxiv-2609.13053", "sourceUrl": "https://arxiv.org/pdf/2609.13053", "sectionsRead": [ "§3.2–3.4", "§5.2–5.7", "Tables 2–3, 5, 9–11", "§6" ], "methodsZh": "在共享掩码扩散主干中序列化文字、视觉和量化动作，联合训练策略、下一图像、任务理解与目标图像预测。动作按块并行迭代解码；默认先生成目标，再条件化动作，也研究联合动作—视觉去噪和固定参考的候选重排。", "experimentsZh": "经OXE机器人继续预训练后，在LIBERO、七类LIBERO-Plus扰动和VLABench两任务评测；Franka Research 3测试水果搬运、分类、自由堆叠及指定颜色顺序堆叠。消融分别固定解码器改变训练目标、固定检查点改变推理组合。", "resultsZh": "作者报告LIBERO平均98.1%、LIBERO-Plus73.0%，真实四条件平均78.4%，颜色顺序堆叠68.5%。完整辅助目标使VLABench域外指令成功由33.88%升至47.28%。B200上块并行解码最高29.15倍加速是模型侧吞吐，非完整机器人周期。", "limitationsZh": "已预留语音/力触觉接口但报告的机器人训练和策略解码未启用，不能称多传感器验证。真实仅单平台单工作区；近未来图像全图指标上复制原帧仍更强。任务理解只定性评估，视觉预测收益取决于推理组合。", "reproductionZh": "区分已发布220k阶段检查点与后续任务模型，记录量化区间、目标混合比例、去噪步数和完整延迟；分别检验任务成功与视觉变化预测。未运行代码或独立复核作者成绩。", "experimentType": "both", "robots": [ "Franka Research 3" ], "evidenceNotes": [ { "section/page": "§3.2.2; §3.4", "note": "传感器接口未启用；目标条件和联合去噪不同推理路径。" }, { "section/page": "Tables 2–3,5,9–11; §6", "note": "仿真/实机成绩、模型侧速度及视觉预测限制。" } ], "corrections": [ { "field": "modalityNote", "value": "实际机器人实验未启用可选语音、力/触觉传感器令牌", "reason": "§3.2.2明确接口容量与已运行实验不同。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04" }, "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2603.09030", "title": "PlayWorld: Learning Robot World Models from Autonomous Play", "shortTitle": "PlayWorld", "date": "2026-03-09", "updatedDate": "2026-04-06", "category": "世界模型 / 自主数据 / 强化学习", "freshness": "2026 较早研究，实作优先", "summary": "PlayWorld用VLM提出场景相关且轻微变化的任务，让现有VLA自主玩耍，采集抓空、滑移、碰撞、变形等成功示范较少覆盖的转移。加关节/工作区安全限制与语言重置，在DROID操作配置上持续采集；从DROID预…", "whyUseful": "先按交互类别检查覆盖及等小时数据对照，再校准预测成功率与真实成功率；课程难度来源、三视角同步和复位策略需保留。策略微调应保留真实验证集与保守停止点，防止想象奖励上涨但真机退化。", "paperUrl": "https://arxiv.org/abs/2603.09030", "projectUrl": "https://robot-playworld.github.io/", "codeUrl": "https://github.com/irom-princeton/open-world", "dataUrl": "https://huggingface.co/datasets/tennyyyin/playworld_dataset_preview", "trainingStatus": "训练代码已核实，部分流程未完成", "trainingNote": "作者项目页指向 open-world；README 列出 bidirectional-SVD、AR-Wan 世界模型训练及多类策略训练。AR-Cosmos 世界模型训练/部分推理仍标 TODO；数据链接名称为 preview。", "caveats": "不是 9–10 月最新论文；公开平台不等同于所有实验资产和全量数据已完整开放。", "url": "https://arxiv.org/abs/2603.09030", "status": "训练代码已核实，部分流程未完成", "titleZh": "PlayWorld：从自主探索交互中学习机器人世界模型", "abstractZh": "PlayWorld用VLM提出场景相关且轻微变化的任务，让现有VLA自主玩耍，采集抓空、滑移、碰撞、变形等成功示范较少覆盖的转移。加关节/工作区安全限制与语言重置，在DROID操作配置上持续采集；从DROID预训练多视角SVD模型出发按难度课程微调。\n三组对象含碗胡萝卜玩偶、积木和毛巾，累计30小时自主数据，与6小时人类成功示范及人类玩耍比较；500多片段、20多个策略组成六类交互评测。用生成世界评估策略，并以DSRL优化冻结扩散策略的初始噪声后回到实机。", "experimentType": "both", "experimentNote": "三组对象含碗胡萝卜玩偶、积木和毛巾，累计30小时自主数据，与6小时人类成功示范及人类玩耍比较；500多片段、20多个策略组成六类交互评测。用生成世界评估策略，并以DSRL优化冻结扩散策略的初始噪声后回到实机。", "robots": [ "DROID操作配置（正文未明确机械臂型号）" ], "robotNote": "论文 §3.2 明确采用无需额外硬件修改的 DROID 操作平台，但已读正文与项目页未直接列出具体机械臂型号，不凭外观补写。", "tags": [ "世界模型", "自主探索", "机器人数据", "策略评估", "模型内强化学习" ], "contribution": "PlayWorld用VLM提出场景相关且轻微变化的任务，让现有VLA自主玩耍，采集抓空、滑移、碰撞、变形等成功示范较少覆盖的转移。加关节/工作区安全限制与语言重置，在DROID操作配置上持续采集；从DROID预训练多视角SVD模型出发按难度课程微调。", "limitations": "依赖已有可执行VLA，采集仍冗余；视频幻觉、动作控制模式差异和长时误差未消除。长期模型内训练可能利用模型漏洞并降低实机表现，规模规律仅在受控实验室和有限对象内验证；原文只称DROID配置，未据此猜商业臂型号。", "codeStatus": "unknown", "codeStatusNote": "官方 open-world 仓库有实际训练与评估源码，但根目录/README 未核实项目级开源许可证；按严格许可口径标未核实，不等于没有代码。", "license": "未核实项目级许可证", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/html/2603.09030v3", "note": "最新 v3（2026-04-06）；§3.2 指向 DROID 平台，§4/§10 为模型内训练与实机测试，§6 给出局限。" }, { "url": "https://robot-playworld.github.io/", "note": "官方项目页链接 open-world 仓库，展示自主交互、策略评估与模型内微调。" }, { "url": "https://github.com/irom-princeton/open-world", "note": "源码目录和训练/评估流程可见；AR-Cosmos 部分功能仍 TODO，未核实项目级许可证。" }, { "url": "https://arxiv.org/pdf/2603.09030", "note": "3.2–3.3; 4.1–4.4：自主数据、DROID配置、交互基准与相关系数。" }, { "url": "https://arxiv.org/pdf/2603.09030", "note": "4.5–4.6; 6; Appendix D：世界内微调后实机验证和长期训练退化。" }, { "url": "https://arxiv.org/pdf/2603.09030v3", "note": "3.2–3.3; 4.1–4.4：自主数据、DROID配置、交互基准与相关系数。" }, { "url": "https://arxiv.org/pdf/2603.09030v3", "note": "4.5–4.6; 6; Appendix D：世界内微调后实机验证和长期训练退化。" } ], "tier": "recent", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对范围为公开原文、平台型号与所列来源；性能数字均为作者报告，未独立复现。", "year": 2026, "directions": [ "世界模型", "安全与评估", "强化学习" ], "robotFilters": [], "original": { "id": "arxiv-2603.09030", "originalSourceUrl": "https://arxiv.org/pdf/2603.09030v3", "versionedOriginalUrl": "https://arxiv.org/pdf/2603.09030v3", "finalDownloadUrl": "https://arxiv.org/pdf/2603.09030", "sha256": "18c8bfc90fad9fe6cc387f923035c823c7061f40cada1eae134b62638e3d187f", "bytes": 11903343, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:33:46.544895+00:00", "publicationDate": "2026/03/09", "publicationDateSourceUrl": "https://arxiv.org/abs/2603.09030", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2603.09030", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2603.09030", "pages": 25, "sourceIdentityStatus": "title_match", "sourceTitle": "PlayWorld: Learning Robot World Models from Autonomous Play", "sourceVersion": "2603.09030v3", "sourceVersionDate": "2026/04/06", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.708726+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2603.09030" }, "originalAnalysis": { "id": "arxiv-2603.09030", "sourceUrl": "https://arxiv.org/pdf/2603.09030v3", "sectionsRead": [ "3.1–3.3", "4.1–4.6", "6", "Appendix D", "Table 1" ], "methodsZh": "PlayWorld用VLM提出场景相关且轻微变化的任务，让现有VLA自主玩耍，采集抓空、滑移、碰撞、变形等成功示范较少覆盖的转移。加关节/工作区安全限制与语言重置，在DROID操作配置上持续采集；从DROID预训练多视角SVD模型出发按难度课程微调。", "experimentsZh": "三组对象含碗胡萝卜玩偶、积木和毛巾，累计30小时自主数据，与6小时人类成功示范及人类玩耍比较；500多片段、20多个策略组成六类交互评测。用生成世界评估策略，并以DSRL优化冻结扩散策略的初始噪声后回到实机。", "resultsZh": "课程后滑移LPIPS0.070，对成功示范训练0.090；预测与真实策略成功率Pearson相关0.8766。两项世界模型内微调实验产生真实行为提升及恢复动作；扩大对象多样性改善未见水果和积木预测，而非仅记住场景。", "limitationsZh": "依赖已有可执行VLA，采集仍冗余；视频幻觉、动作控制模式差异和长时误差未消除。长期模型内训练可能利用模型漏洞并降低实机表现，规模规律仅在受控实验室和有限对象内验证；原文只称DROID配置，未据此猜商业臂型号。", "reproductionZh": "先按交互类别检查覆盖及等小时数据对照，再校准预测成功率与真实成功率；课程难度来源、三视角同步和复位策略需保留。策略微调应保留真实验证集与保守停止点，防止想象奖励上涨但真机退化。", "experimentType": "both", "robots": [ "DROID操作配置（正文未明确机械臂型号）" ], "corrections": [ "both包含学习到的视频模拟器内策略训练与真实机器人测试，不应误写为传统物理仿真结果。" ], "evidenceNotes": [ { "note": "自主数据、DROID配置、交互基准与相关系数。", "section": "3.2–3.3; 4.1–4.4" }, { "note": "世界内微调后实机验证和长期训练退化。", "section": "4.5–4.6; 6; Appendix D" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2603.09030v3", "originalSha256": "18c8bfc90fad9fe6cc387f923035c823c7061f40cada1eae134b62638e3d187f" }, "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2602.06643", "title": "Humanoid Manipulation Interface: Humanoid Whole-Body Manipulation from Robot-Free Demonstrations", "shortTitle": "HuMI", "date": "2026-02-06", "updatedDate": "2026-02-12", "category": "人形全身操作 / 人类示范", "freshness": "2026 较早研究，实作优先", "summary": "HuMI用手持感知夹爪与腰脚追踪器收集无需机器人在场的示范，在线逆运动学预览让人主动避开机器人不可达姿态。高层扩散策略生成身体关键点轨迹，仿真训练的低层全身控制器采用速度自适应末端精度和变速增强；动作段以上一目…", "whyUseful": "复现应共同训练视觉高层和运动跟踪低层，保存示范同步、夹爪坐标及关键点相对参考。高层5Hz、低层50Hz；预览与参考选择消融有大幅下降，单纯复制网络架构不足以复现整体系统。", "paperUrl": "https://arxiv.org/abs/2602.06643", "projectUrl": "https://humanoid-manipulation-interface.github.io/", "codeUrl": "https://github.com/Richard-coder-Nai/HuMI", "trainingStatus": "完整训练入口已核实", "trainingNote": "官方仓库有采集、处理、高层训练、低层训练、部署五部分。低层基于 Isaac Lab、cuRobo、PPO teacher/student；高层提供 train.sh 与加速训练配置。", "caveats": "不是 9–10 月最新论文；依赖 G1 平台、示范动作数据和采集硬件，实机部署仍需校准与安全验证。", "extraSources": [ "https://github.com/Richard-coder-Nai/HuMI/blob/main/humi_high_level_policy/README.md", "https://github.com/Richard-coder-Nai/HuMI/blob/main/humi_low_level_controller/README.md" ], "url": "https://arxiv.org/abs/2602.06643", "status": "完整训练入口已核实", "titleZh": "人形机器人操作接口：从无需机器人在场的示范中学习人形机器人全身操作", "abstractZh": "HuMI用手持感知夹爪与腰脚追踪器收集无需机器人在场的示范，在线逆运动学预览让人主动避开机器人不可达姿态。高层扩散策略生成身体关键点轨迹，仿真训练的低层全身控制器采用速度自适应末端精度和变速增强；动作段以上一目标为参考，腰脚等缺乏视觉锚点处只跟踪段内相对运动。\nUnitree G1实机评跪地取环、双臂抽剑、投掷及走向清桌，各20次；另用七环境、七瓶型350示范训练蹲取策略，在四新环境及六新物体共20次测试。15分钟采集对照TWIST2，均由有经验操作者执行。", "experimentType": "both", "experimentNote": "Unitree G1实机评跪地取环、双臂抽剑、投掷及走向清桌，各20次；另用七环境、七瓶型350示范训练蹲取策略，在四新环境及六新物体共20次测试。15分钟采集对照TWIST2，均由有经验操作者执行。", "robots": [ "Unitree G1" ], "robotNote": "实机使用定制 UMI 风格夹爪和 HTC Vive Ultimate Tracker 定位。", "tags": [ "人形机器人", "全身操作", "人类示范", "模仿学习", "层级控制", "Diffusion Policy" ], "contribution": "HuMI用手持感知夹爪与腰脚追踪器收集无需机器人在场的示范，在线逆运动学预览让人主动避开机器人不可达姿态。高层扩散策略生成身体关键点轨迹，仿真训练的低层全身控制器采用速度自适应末端精度和变速增强；动作段以上一目标为参考，腰脚等缺乏视觉锚点处只跟踪段内相对运动。", "limitations": "机器人不在采集现场仍需追踪器、标定及人看IK预览，不是任意自然视频直接学习。追踪依赖纹理照明，低层控制器尚非通用，只有G1验证；多数能力测试与示范同场景，70%泛化仅对应蹲取任务。", "codeStatus": "open", "codeStatusNote": "官方仓库有采集、训练、部署五部分实现及 MIT 许可；未实际执行完整训练。", "license": "MIT", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/html/2602.06643v2", "note": "§II-B 说明仿真训练；§III-VI 及附录 B 明确 G1 实机评估；§VIII 说明局限。" }, { "url": "https://github.com/Richard-coder-Nai/HuMI", "note": "仓库列出采集、处理、高层训练、低层训练和部署完整入口。" }, { "url": "https://github.com/Richard-coder-Nai/HuMI/blob/main/LICENSE", "note": "可读 MIT 许可证。" }, { "url": "https://arxiv.org/pdf/2602.06643v2", "note": "II-A–C：原始未缩放人体轨迹配在线IK；分层参考接口是关键。" }, { "url": "https://arxiv.org/pdf/2602.06643v2", "note": "IV–V：域内四任务20次各；70%来自蹲取泛化14/20。" }, { "url": "https://arxiv.org/pdf/2602.06643v2", "note": "VI; VIII：15分钟抽剑采集效率；视觉追踪、单平台、非通用低层限制。" } ], "tier": "recent", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对范围为公开原文、平台型号与所列来源；性能数字均为作者报告，未独立复现。", "year": 2026, "directions": [ "运动控制", "操作与抓取", "人形机器人", "模仿学习" ], "robotFilters": [ "Unitree G1" ], "original": { "id": "arxiv-2602.06643", "originalSourceUrl": "https://arxiv.org/pdf/2602.06643v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2602.06643v2", "finalDownloadUrl": "https://arxiv.org/pdf/2602.06643", "sha256": "7d73380b98fccd23e69206db2ec0d875634396c62fba24ee277b6679fd2f5e45", "bytes": 37246617, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:33:51.460287+00:00", "publicationDate": "2026/02/06", "publicationDateSourceUrl": "https://arxiv.org/abs/2602.06643", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2602.06643", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2602.06643", "pages": 15, "sourceIdentityStatus": "title_match", "sourceTitle": "Humanoid Manipulation Interface: Humanoid Whole-Body Manipulation from Robot-Free Demonstrations", "sourceVersion": "2602.06643v2", "sourceVersionDate": "2026/02/12", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.756134+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2602.06643" }, "originalAnalysis": { "id": "arxiv-2602.06643", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458243+00:00", "sourceUrl": "https://arxiv.org/pdf/2602.06643v2", "sectionsRead": [ "II-A–C", "III–VI Experiments", "VIII Limitations", "Appendix E–F", "Downloaded latest-version methods/results cross-check" ], "methodsZh": "HuMI用手持感知夹爪与腰脚追踪器收集无需机器人在场的示范，在线逆运动学预览让人主动避开机器人不可达姿态。高层扩散策略生成身体关键点轨迹，仿真训练的低层全身控制器采用速度自适应末端精度和变速增强；动作段以上一目标为参考，腰脚等缺乏视觉锚点处只跟踪段内相对运动。", "experimentsZh": "Unitree G1实机评跪地取环、双臂抽剑、投掷及走向清桌，各20次；另用七环境、七瓶型350示范训练蹲取策略，在四新环境及六新物体共20次测试。15分钟采集对照TWIST2，均由有经验操作者执行。", "resultsZh": "作者报告四项域内成功率85%、85%、75%、75%；新环境/物体蹲取14/20即70%。抽剑示范采集62段对28段，接受率96.7%对64.3%，单个可用示范耗时约为对照30%；“三倍效率”对应这一特定流程。", "limitationsZh": "机器人不在采集现场仍需追踪器、标定及人看IK预览，不是任意自然视频直接学习。追踪依赖纹理照明，低层控制器尚非通用，只有G1验证；多数能力测试与示范同场景，70%泛化仅对应蹲取任务。", "reproductionZh": "复现应共同训练视觉高层和运动跟踪低层，保存示范同步、夹爪坐标及关键点相对参考。高层5Hz、低层50Hz；预览与参考选择消融有大幅下降，单纯复制网络架构不足以复现整体系统。", "experimentType": "both", "robots": [ "Unitree G1" ], "evidenceNotes": [ { "section": "II-A–C", "note": "原始未缩放人体轨迹配在线IK；分层参考接口是关键。" }, { "section": "IV–V", "note": "域内四任务20次各；70%来自蹲取泛化14/20。" }, { "section": "VI; VIII", "note": "15分钟抽剑采集效率；视觉追踪、单平台、非通用低层限制。" } ], "corrections": [ "70%不可泛化为五类任务共同跨环境成功率；实际仅蹲取评测。" ], "sourceVersion": "2602.06643v2", "originalSha256": "7d73380b98fccd23e69206db2ec0d875634396c62fba24ee277b6679fd2f5e45", "additionalSourcesRead": [ "https://arxiv.org/html/2602.06643v1" ] }, "analysisVerifiedAt": "2026-10-04T13:45:34.458243+00:00" }, { "id": "arxiv-2511.14759", "title": "π*0.6: a VLA That Learns From Experience", "titleZh": "π*0.6：从经验中学习的视觉—语言—动作模型", "shortTitle": "π*0.6 / RECAP", "date": "2025-11-18", "year": 2025, "url": "https://arxiv.org/abs/2511.14759", "paperUrl": "https://arxiv.org/abs/2511.14759", "projectUrl": "https://pi.website/blog/pistar06", "codeUrl": null, "summary": "Recap用回报分布价值预测估计数据中动作的优势，将二值改进条件加入流匹配VLA；混合原示范、自主执行与专家接管数据，重复收集一批再离线更新。π*0.6以Gemma3 4B及860M动作专家生成动作块，不靠每次…", "abstractZh": "Recap用回报分布价值预测估计数据中动作的优势，将二值改进条件加入流匹配VLA；混合原示范、自主执行与专家接管数据，重复收集一批再离线更新。π*0.6以Gemma3 4B及860M动作专家生成动作块，不靠每次任务在线PPO更新。\n定量测试衣物折叠、双份浓缩咖啡和纸箱组装，使用两台六自由度机械臂的固定双臂平台和三相机。不同任务分别设200/500/600秒成功时限；叠衣每迭代300自主回合，纸箱600自主加360接管回合。", "category": "VLA / 实机强化学习", "tags": [ "π*0.6 / RECAP", "视觉语言动作", "模仿学习", "操作与抓取" ], "directions": [ "视觉语言动作", "操作与抓取", "强化学习", "模仿学习" ], "tier": "recent", "experimentType": "real", "experimentNote": "定量测试衣物折叠、双份浓缩咖啡和纸箱组装，使用两台六自由度机械臂的固定双臂平台和三相机。不同任务分别设200/500/600秒成功时限；叠衣每迭代300自主回合，纸箱600自主加360接管回合。", "robots": [ "固定双臂平台（两台6自由度机械臂，型号未给出）" ], "robotFilters": [], "robotNote": "原文图5仅明确静态双臂、每臂6自由度与平行夹爪；厂商和型号未知。", "codeStatus": "unknown", "status": "官方代码未核实", "trainingStatus": "官方代码未核实", "trainingNote": "论文与模型卡可查；不能把其他 π 系列的 openpi 代码直接当作该版本 RECAP 的完整训练实现，状态保持未知。", "codeStatusNote": "本轮未核实可复现该论文的官方训练仓库；不据此断言从未发布。", "license": "未核实", "contribution": "Recap用回报分布价值预测估计数据中动作的优势，将二值改进条件加入流匹配VLA；混合原示范、自主执行与专家接管数据，重复收集一批再离线更新。π*0.6以Gemma3 4B及860M动作专家生成动作块，不靠每次任务在线PPO更新。", "whyUseful": "复现需明确每版本预训练/任务数据、优势标注、人工评分和超时规则，吞吐必须同时包含速度与成功。论文仅称固定双6DoF平台，不能从照片猜厂商；本轮未核验全量训练数据及端到端复现入口。", "limitations": "复杂衣物训练含11类，但主要定量指标针对衬衫；咖啡指标也只针对双份浓缩，不能扩成所有饮品。系统依赖人工奖励、接管和场景重置；更新为批式离线，探索较简单，尚非完全自主持续在线学习。", "caveats": "依赖人工标注、干预和复位；探索较保守；非完全在线学习，完整训练代码未核实。", "evidence": [ { "url": "https://arxiv.org/abs/2511.14759", "note": "首发 2025-11-18；正式论文确实存在，早于本轮核对日期。" }, { "url": "https://arxiv.org/html/2511.14759v1", "note": "图5：双 6-DoF 静态双臂；§VII：人工监督、复位及批量离线迭代限制。" }, { "url": "https://website.pi-asset.com/pi06star/PI06_model_card.pdf", "note": "官方 π0.6 基础模型卡日期 2025-11-17；与 π*0.6 强化学习版本区分。" }, { "url": "https://arxiv.org/html/2511.14759v1", "note": "Fig.5：固定双6DoF平行夹爪平台，未提供厂商型号。" }, { "url": "https://arxiv.org/html/2511.14759v1", "note": "§VI-A：衣物/咖啡定量任务范围与各自时限。" }, { "url": "https://arxiv.org/html/2511.14759v1", "note": "§VII：人工奖励与重置、批式离线更新。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2025：将真实部署、自主尝试与人工纠错纳入大型 VLA 的迭代强化学习。", "freshness": "近年进展", "original": { "id": "arxiv-2511.14759", "originalSourceUrl": "https://arxiv.org/pdf/2511.14759v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2511.14759v2", "finalDownloadUrl": "https://arxiv.org/pdf/2511.14759", "sha256": "9f4f5a3070ad9549ea7687ea54baa19858820dbcc0834cc805212d68fd6c8cb3", "bytes": 25075375, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:11.755411+00:00", "publicationDate": "2025/11/18", "publicationDateSourceUrl": "https://arxiv.org/abs/2511.14759", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2511.14759", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2511.14759", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "$\\pi^{*}_{0.6}$: a VLA That Learns From Experience", "sourceVersion": "2511.14759v2", "sourceVersionDate": "2025/11/19", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.780675+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2511.14759" }, "originalAnalysis": { "id": "arxiv-2511.14759", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090127+00:00", "sourceUrl": "https://arxiv.org/html/2511.14759v1", "sectionsRead": [ "§IV", "§V", "§VI-A–C", "Figs.5,7–12", "§VII" ], "methodsZh": "Recap用回报分布价值预测估计数据中动作的优势，将二值改进条件加入流匹配VLA；混合原示范、自主执行与专家接管数据，重复收集一批再离线更新。π*0.6以Gemma3 4B及860M动作专家生成动作块，不靠每次任务在线PPO更新。", "experimentsZh": "定量测试衣物折叠、双份浓缩咖啡和纸箱组装，使用两台六自由度机械臂的固定双臂平台和三相机。不同任务分别设200/500/600秒成功时限；叠衣每迭代300自主回合，纸箱600自主加360接管回合。", "resultsZh": "作者报告复杂衣物与咖啡的每小时成功吞吐较离线RL加SFT超过翻倍、失败率约减半；除复杂衣物外最终成功率超过90%。固定橙色T恤特定失败模式用每轮600轨迹、两轮后达97%。", "limitationsZh": "复杂衣物训练含11类，但主要定量指标针对衬衫；咖啡指标也只针对双份浓缩，不能扩成所有饮品。系统依赖人工奖励、接管和场景重置；更新为批式离线，探索较简单，尚非完全自主持续在线学习。", "reproductionZh": "复现需明确每版本预训练/任务数据、优势标注、人工评分和超时规则，吞吐必须同时包含速度与成功。论文仅称固定双6DoF平台，不能从照片猜厂商；本轮未核验全量训练数据及端到端复现入口。", "experimentType": "real", "robots": [ "固定双臂平台（两台6自由度机械臂，型号未给出）" ], "evidenceNotes": [ { "section/page": "Fig.5", "note": "固定双6DoF平行夹爪平台，未提供厂商型号。" }, { "section/page": "§VI-A", "note": "衣物/咖啡定量任务范围与各自时限。" }, { "section/page": "§VII", "note": "人工奖励与重置、批式离线更新。" } ], "corrections": { "trainingMode": "迭代式离线RL加实机数据收集；非同时实时更新的在线RL。" } }, "analysisVerifiedAt": "2026-10-04T13:50:35.090127+00:00" }, { "id": "arxiv-2506.09985", "title": "V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning", "titleZh": "V-JEPA 2：自监督视频模型实现理解、预测与规划", "shortTitle": "V-JEPA 2", "date": "2025-06-11", "year": 2025, "url": "https://arxiv.org/abs/2506.09985", "paperUrl": "https://arxiv.org/abs/2506.09985", "projectUrl": "https://ai.meta.com/research/vjepa/", "codeUrl": "https://github.com/facebookresearch/vjepa2", "summary": "先以视频遮挡表征预测学习编码器，再冻结编码器，用DROID视频及末端状态训练动作条件自回归潜在预测器，结合教师强制和两步展开损失。控制时用CEM最小化预测潜在状态与目标图像的L1距离，每次仅执行首动作再规划。", "abstractZh": "先以视频遮挡表征预测学习编码器，再冻结编码器，用DROID视频及末端状态训练动作条件自回归潜在预测器，结合教师强制和两步展开损失。控制时用CEM最小化预测潜在状态与目标图像的L1距离，每次仅执行首动作再规划。\n动作条件后训练使用不足62小时DROID片段，不需要任务奖励或成功标签但需要机器人状态。相同模型在两个未见实验室的Franka Panda与RobotiQ夹爪上评测到达、抓取及搬运，每技能条件10次；并比较微调Octo与Cosmos。", "category": "世界模型 / 潜在视觉规划", "tags": [ "V-JEPA 2", "世界模型", "强化学习", "操作与抓取" ], "directions": [ "世界模型", "操作与抓取", "强化学习" ], "tier": "recent", "experimentType": "real", "experimentNote": "动作条件后训练使用不足62小时DROID片段，不需要任务奖励或成功标签但需要机器人状态。相同模型在两个未见实验室的Franka Panda与RobotiQ夹爪上评测到达、抓取及搬运，每技能条件10次；并比较微调Octo与Cosmos。", "robots": [ "Franka Emika Panda", "RobotiQ夹爪" ], "robotFilters": [ "Franka Panda", "RobotiQ夹爪" ], "robotNote": "两个实验室使用 Franka Emika Panda 配 RobotiQ 夹爪；夹爪更细型号未核实。", "codeStatus": "open", "status": "训练代码已开源", "trainingStatus": "训练代码已开源", "trainingNote": "官方仓库提供 V-JEPA 2 和 2-AC 代码、配置与模型入口；多数代码 MIT，部分文件 Apache-2.0。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "MIT 为主，部分代码 Apache-2.0", "contribution": "先以视频遮挡表征预测学习编码器，再冻结编码器，用DROID视频及末端状态训练动作条件自回归潜在预测器，结合教师强制和两步展开损失。控制时用CEM最小化预测潜在状态与目标图像的L1距离，每次仅执行首动作再规划。", "whyUseful": "应固定DROID筛选、状态差分动作、图像目标与子目标切换规则，同时报告候选数、时域和总控制延迟；将视频理解基准与机器人闭环成绩分开。未运行模型或实机。", "limitations": "搬运使用中间目标图像，零样本指新实验室部署，不代表未经机器人数据训练。长时域模型误差与搜索成本增加，现有控制目标为图像；小规模物体评测无法证明开放世界操作能力。", "caveats": "相机位置敏感；长期预测与搜索会累积误差；需要图像目标，规划时延仍明显。", "evidence": [ { "url": "https://arxiv.org/abs/2506.09985", "note": "首发日期。" }, { "url": "https://arxiv.org/html/2506.09985v1", "note": "§3、§4：DROID 后训练、Panda 实机及相机/规划/目标限制。" }, { "url": "https://github.com/facebookresearch/vjepa2", "note": "官方 2/2-AC 实现及分组件许可说明。" }, { "url": "https://arxiv.org/pdf/2506.09985", "note": "§3.1; §4.1：不足62小时训练输入与两实验室Panda部署。" }, { "url": "https://arxiv.org/pdf/2506.09985", "note": "Tables 2–3; Appendix B.2：每条件10次、规划16秒、搬运含子目标。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2025：大规模自监督视频表征结合小规模机器人动作后训练，实现新环境实机规划。", "freshness": "近年进展", "original": { "id": "arxiv-2506.09985", "originalSourceUrl": "https://arxiv.org/pdf/2506.09985v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2506.09985v1", "finalDownloadUrl": "https://arxiv.org/pdf/2506.09985", "sha256": "9cfcfde5fb0d9730637da5b9e7317825c3f3d09e91f3553e22eeba42c74d2226", "bytes": 20240380, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:16.510702+00:00", "publicationDate": "2025/06/11", "publicationDateSourceUrl": "https://arxiv.org/abs/2506.09985", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2506.09985", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2506.09985", "pages": 48, "sourceIdentityStatus": "title_match", "sourceTitle": "V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning", "sourceVersion": "2506.09985v1", "sourceVersionDate": "2025/06/11", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.800205+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2506.09985" }, "originalAnalysis": { "id": "arxiv-2506.09985", "sourceUrl": "https://arxiv.org/pdf/2506.09985", "sectionsRead": [ "§3.1–3.2", "§4.1–4.3", "Tables 2–3", "§9" ], "methodsZh": "先以视频遮挡表征预测学习编码器，再冻结编码器，用DROID视频及末端状态训练动作条件自回归潜在预测器，结合教师强制和两步展开损失。控制时用CEM最小化预测潜在状态与目标图像的L1距离，每次仅执行首动作再规划。", "experimentsZh": "动作条件后训练使用不足62小时DROID片段，不需要任务奖励或成功标签但需要机器人状态。相同模型在两个未见实验室的Franka Panda与RobotiQ夹爪上评测到达、抓取及搬运，每技能条件10次；并比较微调Octo与Cosmos。", "resultsZh": "作者报告两实验室平均杯/盒抓取65%/25%，杯/盒搬运80%/65%；结果依任务明显变化。RTX4090上使用800候选、十轮、一步时域，规划每动作16秒；Cosmos比较为80候选、每动作约4分钟，不能称高频实时控制。", "limitationsZh": "搬运使用中间目标图像，零样本指新实验室部署，不代表未经机器人数据训练。长时域模型误差与搜索成本增加，现有控制目标为图像；小规模物体评测无法证明开放世界操作能力。", "reproductionZh": "应固定DROID筛选、状态差分动作、图像目标与子目标切换规则，同时报告候选数、时域和总控制延迟；将视频理解基准与机器人闭环成绩分开。未运行模型或实机。", "experimentType": "real", "robots": [ "Franka Emika Panda", "RobotiQ夹爪" ], "evidenceNotes": [ { "section/page": "§3.1; §4.1", "note": "不足62小时训练输入与两实验室Panda部署。" }, { "section/page": "Tables 2–3; Appendix B.2", "note": "每条件10次、规划16秒、搬运含子目标。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2506.01844", "title": "SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics", "titleZh": "SmolVLA：面向经济高效机器人的视觉—语言—动作模型", "shortTitle": "SmolVLA", "date": "2025-06-02", "year": 2025, "url": "https://arxiv.org/abs/2506.01844", "paperUrl": "https://arxiv.org/abs/2506.01844", "projectUrl": null, "codeUrl": "https://github.com/huggingface/lerobot", "summary": "SmolVLA冻结紧凑SmolVLM-2，以视觉/语言/状态特征条件化流匹配动作专家；图像压到每帧64词元，只用前16层，并交替交叉注意力和因果自注意力。另提出让动作执行与推理解耦的异步管线，以减少动作块之间等…", "abstractZh": "SmolVLA冻结紧凑SmolVLM-2，以视觉/语言/状态特征条件化流匹配动作专家；图像压到每帧64词元，只用前16层，并交替交叉注意力和因果自注意力。另提出让动作执行与推理解耦的异步管线，以减少动作块之间等待。\nLIBERO40任务与Meta-World50任务各每任务10次；SO100抓放、堆叠、分类及SO101抓放。实机预训练约2.3万社区轨迹，仿真表2的SmolVLA却仅从VLM初始化，需区别训练来源。", "category": "VLA / 低成本机器人", "tags": [ "SmolVLA", "视觉语言动作", "模仿学习", "操作与抓取" ], "directions": [ "视觉语言动作", "操作与抓取", "模仿学习" ], "tier": "recent", "experimentType": "both", "experimentNote": "LIBERO40任务与Meta-World50任务各每任务10次；SO100抓放、堆叠、分类及SO101抓放。实机预训练约2.3万社区轨迹，仿真表2的SmolVLA却仅从VLM初始化，需区别训练来源。", "robots": [ "SO100", "SO101", "Franka Emika Panda（仿真）", "Sawyer（仿真）" ], "robotFilters": [ "SO-100", "SO-101", "Franka Panda", "Sawyer" ], "robotNote": "SO100、SO101 为实机；Panda 与 Sawyer 分别用于 LIBERO 和 Meta-World 仿真。", "codeStatus": "open", "status": "训练代码已开源", "trainingStatus": "训练代码已开源", "trainingNote": "LeRobot 提供 SmolVLA 训练和推理代码，Apache-2.0；权重与社区数据应逐项核对许可。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "Apache-2.0（代码；权重和数据单独核对）", "contribution": "SmolVLA冻结紧凑SmolVLM-2，以视觉/语言/状态特征条件化流匹配动作专家；图像压到每帧64词元，只用前16层，并交替交叉注意力和因果自注意力。另提出让动作执行与推理解耦的异步管线，以减少动作块之间等待。", "whyUseful": "以LeRobot实现固定50步动作块、10步流求解和冻结VLM开始；区分同步主表与异步实验，且仿真每步重观测、实机整块后观测。重跑时同时给完整任务成功率和论文部分得分。", "limitations": "机器人社区数据形态和任务较集中，长时序能力未确立；不同基线训练数据量不同，不能直接归因参数效率。小模型可单卡训练不意味着全文实验便宜，作者报告整个项目约3万GPU小时。", "caveats": "预训练主要来自 SO100；任务多样性和跨本体能力有限，异步执行仍需处理延迟和旧观察。", "evidence": [ { "url": "https://arxiv.org/abs/2506.01844", "note": "首发日期和正式标题。" }, { "url": "https://arxiv.org/html/2506.01844v1", "note": "§4.2：SO100/SO101 实机、Panda/Sawyer 仿真；§5.1 数据与本体局限。" }, { "url": "https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/smolvla/modeling_smolvla.py", "note": "官方模型实现，文件声明 Apache-2.0。" }, { "url": "https://arxiv.org/pdf/2506.01844", "note": "3.1; 4.1; 4.3：紧凑架构、实机部分计分及推理协议。" }, { "url": "https://arxiv.org/pdf/2506.01844", "note": "Tables 2–5; 5.1：训练来源差异、成绩与数据/任务边界。" }, { "url": "https://arxiv.org/pdf/2506.01844v1", "note": "3.1; 4.1; 4.3：紧凑架构、实机部分计分及推理协议。" }, { "url": "https://arxiv.org/pdf/2506.01844v1", "note": "Tables 2–5; 5.1：训练来源差异、成绩与数据/任务边界。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2025：轻量化 VLA 与社区低成本硬件数据降低复现和部署门槛。", "freshness": "近年进展", "original": { "id": "arxiv-2506.01844", "originalSourceUrl": "https://arxiv.org/pdf/2506.01844v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2506.01844v1", "finalDownloadUrl": "https://arxiv.org/pdf/2506.01844", "sha256": "cb1bb9a8f824187fcdc32af8c290214487c524f139a3c2def1c6d97adb1fa40c", "bytes": 8921356, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:21.628286+00:00", "publicationDate": "2025/06/02", "publicationDateSourceUrl": "https://arxiv.org/abs/2506.01844", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2506.01844", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2506.01844", "pages": 24, "sourceIdentityStatus": "title_match", "sourceTitle": "SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics", "sourceVersion": "2506.01844v1", "sourceVersionDate": "2025/06/02", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.809190+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2506.01844" }, "originalAnalysis": { "id": "arxiv-2506.01844", "sourceUrl": "https://arxiv.org/pdf/2506.01844v1", "sectionsRead": [ "3.1–3.3", "4.1–4.6", "5.1", "Tables 2–5" ], "methodsZh": "SmolVLA冻结紧凑SmolVLM-2，以视觉/语言/状态特征条件化流匹配动作专家；图像压到每帧64词元，只用前16层，并交替交叉注意力和因果自注意力。另提出让动作执行与推理解耦的异步管线，以减少动作块之间等待。", "experimentsZh": "LIBERO40任务与Meta-World50任务各每任务10次；SO100抓放、堆叠、分类及SO101抓放。实机预训练约2.3万社区轨迹，仿真表2的SmolVLA却仅从VLM初始化，需区别训练来源。", "resultsZh": "4.5亿参数版LIBERO87.3%、Meta-World57.3%；SO100平均78.3，对π0的61.7和ACT48.3。预训练把同构多任务版本51.7提升到78.3；SO101分布内90、位置外推50。实机“成功率”实际按抓取/放置各0.5计部分完成。", "limitationsZh": "机器人社区数据形态和任务较集中，长时序能力未确立；不同基线训练数据量不同，不能直接归因参数效率。小模型可单卡训练不意味着全文实验便宜，作者报告整个项目约3万GPU小时。", "reproductionZh": "以LeRobot实现固定50步动作块、10步流求解和冻结VLM开始；区分同步主表与异步实验，且仿真每步重观测、实机整块后观测。重跑时同时给完整任务成功率和论文部分得分。", "experimentType": "both", "robots": [ "SO100", "SO101", "Franka Emika Panda（仿真）", "Sawyer（仿真）" ], "corrections": [ "实机表中百分比为子任务部分完成得分，不能一概视为二元成功率。" ], "evidenceNotes": [ { "note": "紧凑架构、实机部分计分及推理协议。", "section": "3.1; 4.1; 4.3" }, { "note": "训练来源差异、成绩与数据/任务边界。", "section": "Tables 2–5; 5.1" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2506.01844v1", "originalSha256": "cb1bb9a8f824187fcdc32af8c290214487c524f139a3c2def1c6d97adb1fa40c" }, "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2504.16054", "title": "π0.5: a Vision-Language-Action Model with Open-World Generalization", "titleZh": "π0.5：具有开放世界泛化能力的视觉－语言－动作模型", "shortTitle": "π0.5", "date": "2025-04-22", "year": 2025, "url": "https://arxiv.org/abs/2504.16054", "codeUrl": "https://github.com/Physical-Intelligence/openpi", "category": "视觉语言动作", "tags": [ "开放世界泛化", "移动操作", "异构共训练", "层级策略" ], "tier": "recent", "summary": "π0.5把多场景移动/固定机械臂、跨形态动作、语言子任务和网络视觉数据共同训练。先以FAST离散动作做自回归预训练，再增加流匹配动作专家；同一模型先预测语义子任务，再条件化生成连续动作段，兼顾语言监督与实时控制…", "abstractZh": "π0.5把多场景移动/固定机械臂、跨形态动作、语言子任务和网络视觉数据共同训练。先以FAST离散动作做自回归预训练，再增加流匹配动作专家；同一模型先预测语义子任务，再条件化生成连续动作段，兼顾语言监督与实时控制。\n使用两类自研轮式双臂升降平台，末端/底盘等共18或19维；所有测试场景均未参与训练。三真实住宅的厨房与卧室任务每任务环境十次，另在实体模拟家居场地比较3至104训练地点规模及数据/高层推理消融。约400小时移动示范只是总数据的一部分。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "real", "robots": [ "两类轮式双臂移动操作平台（商业型号未明确）" ], "codeStatus": "open", "status": "openpi 已公开 π0.5 模型与实现；公开版本范围有限。", "trainingNote": "Apache-2.0 与 Gemma 条款；当前实现仅支持 π0.5 流匹配头，公开版本使用 knowledge insulation，不能视为原文全流程原样复现。", "whyUseful": "复现需获得完整混合数据与子任务标注，并区分预训练280K步和后训练80K步；动作专家十步去噪，50Hz指令由分块实现。评估应使用附录进度评分、留出住宅与数据混合消融，不能只引用视频成功片段。", "contribution": "π0.5把多场景移动/固定机械臂、跨形态动作、语言子任务和网络视觉数据共同训练。先以FAST离散动作做自回归预训练，再增加流匹配动作专家；同一模型先预测语义子任务，再条件化生成连续动作段，兼顾语言监督与实时控制。", "limitations": "陌生把手、难开柜门、遮挡和重复开关抽屉仍失败；短上下文及简单指令限制跨房间记忆与复杂偏好。控制链未加独立规划或碰撞检测，研究展示不意味着适合无人监督家居使用，论文未给可确认的商业整机型号。", "caveats": "原文仅描述两类硬件结构，未可靠披露商品型号；不据照片猜测。", "evidence": [ { "url": "https://arxiv.org/abs/2504.16054", "note": "arXiv 首发 2025-04-22。" }, { "url": "https://arxiv.org/html/2504.16054v1", "note": "第 IV-E 节两类平台；第 V–VI 节未见家庭实验与局限。" }, { "url": "https://github.com/Physical-Intelligence/openpi", "note": "公开 π0.5 及流匹配头、版本配方和许可证说明。" }, { "url": "https://arxiv.org/html/2504.16054v1", "note": "IV-C–E：约400小时移动数据外还有其他机器人及网络数据；两类18/19维平台。" }, { "url": "https://arxiv.org/html/2504.16054v1", "note": "V-A; Figure 7, PDF p8：已视觉核对纵轴为Average task progress；每任务/环境十次。" }, { "url": "https://arxiv.org/html/2504.16054v1", "note": "V-C; VI：网络数据的作用依任务而异；把手、遮挡、记忆限制。" } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "视觉语言动作", "操作与抓取" ], "robotFilters": [], "verificationNote": "已核论文与当前官方仓库；明确区分原始论文系统和后续开放实现。", "original": { "id": "arxiv-2504.16054", "originalSourceUrl": "https://arxiv.org/pdf/2504.16054v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2504.16054v1", "finalDownloadUrl": "https://arxiv.org/pdf/2504.16054", "sha256": "6a1029fd8ab6944b74cf22f5e5d30e60bc15699d964b2900af799b807a34b64c", "bytes": 16213397, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:21.915829+00:00", "publicationDate": "2025/04/22", "publicationDateSourceUrl": "https://arxiv.org/abs/2504.16054", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2504.16054", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2504.16054", "pages": 19, "sourceIdentityStatus": "title_match", "sourceTitle": "$\\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization", "sourceVersion": "2504.16054v1", "sourceVersionDate": "2025/04/22", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.822284+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2504.16054" }, "originalAnalysis": { "id": "arxiv-2504.16054", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458245+00:00", "sourceUrl": "https://arxiv.org/html/2504.16054v1", "sectionsRead": [ "IV-A–E", "V-A–E", "Figures 7–13", "VI Discussion", "Appendix A-B–D" ], "methodsZh": "π0.5把多场景移动/固定机械臂、跨形态动作、语言子任务和网络视觉数据共同训练。先以FAST离散动作做自回归预训练，再增加流匹配动作专家；同一模型先预测语义子任务，再条件化生成连续动作段，兼顾语言监督与实时控制。", "experimentsZh": "使用两类自研轮式双臂升降平台，末端/底盘等共18或19维；所有测试场景均未参与训练。三真实住宅的厨房与卧室任务每任务环境十次，另在实体模拟家居场地比较3至104训练地点规模及数据/高层推理消融。约400小时移动示范只是总数据的一部分。", "resultsZh": "作者展示2–5分钟多阶段整理；图7的纵轴是完成进度，实家各任务约65%–95%，不能当作整段二元成功率。增加训练地点及跨形态/跨环境数据提高表现；网络数据在普通整理消融不显著，却帮助新物体指令与高层选择。", "limitationsZh": "陌生把手、难开柜门、遮挡和重复开关抽屉仍失败；短上下文及简单指令限制跨房间记忆与复杂偏好。控制链未加独立规划或碰撞检测，研究展示不意味着适合无人监督家居使用，论文未给可确认的商业整机型号。", "reproductionZh": "复现需获得完整混合数据与子任务标注，并区分预训练280K步和后训练80K步；动作专家十步去噪，50Hz指令由分块实现。评估应使用附录进度评分、留出住宅与数据混合消融，不能只引用视频成功片段。", "experimentType": "real", "robots": [ "两类轮式双臂移动操作平台（商业型号未明确）" ], "evidenceNotes": [ { "section": "IV-C–E", "note": "约400小时移动数据外还有其他机器人及网络数据；两类18/19维平台。" }, { "section": "V-A; Figure 7, PDF p8", "note": "已视觉核对纵轴为Average task progress；每任务/环境十次。" }, { "section": "V-C; VI", "note": "网络数据的作用依任务而异；把手、遮挡、记忆限制。" } ], "corrections": [ "mock homes为实体实验场景，不是计算机仿真；图7分数是任务进度。" ], "sourceVersion": "2504.16054v1", "originalSha256": "6a1029fd8ab6944b74cf22f5e5d30e60bc15699d964b2900af799b807a34b64c" }, "experimentNote": "使用两类自研轮式双臂升降平台，末端/底盘等共18或19维；所有测试场景均未参与训练。三真实住宅的厨房与卧室任务每任务环境十次，另在实体模拟家居场地比较3至104训练地点规模及数据/高层推理消融。约400小时移动示范只是总数据的一部分。", "analysisVerifiedAt": "2026-10-04T13:45:34.458245+00:00" }, { "id": "arxiv-2503.14734", "title": "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots", "titleZh": "GR00T N1：面向通用人形机器人的开放基础模型", "shortTitle": "GR00T N1", "date": "2025-03-18", "year": 2025, "url": "https://arxiv.org/abs/2503.14734", "codeUrl": "https://github.com/NVIDIA/Isaac-GR00T", "category": "机器人基础模型", "tags": [ "人形机器人", "双系统", "合成数据", "流匹配" ], "tier": "recent", "summary": "GR00T N1将Eagle-2视觉语言模块与流匹配Diffusion Transformer相连，以本体专用编码/解码适配不同状态动作维度。混合真实遥操作、物理仿真、人类视频和生成视频，借潜在动作或逆动力学伪…", "abstractZh": "GR00T N1将Eagle-2视觉语言模块与流匹配Diffusion Transformer相连，以本体专用编码/解码适配不同状态动作维度。混合真实遥操作、物理仿真、人类视频和生成视频，借潜在动作或逆动力学伪标签利用无动作视频，再进行任务后训练。\n仿真覆盖RoboCasa24、DexMimicGen9及GR-1桌面24任务，每任务30/100/300示范；真实Fourier GR-1涵盖抓放、关节物体、工业和协作任务。多数实机任务10次，Pack Machinery五次并按物体完成数计分，使用阶段性部分得分。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "Fourier GR-1" ], "codeStatus": "open", "status": "官方代码与 GR00T-N1-2B 权重公开；主仓库已演进到后续版本。", "trainingNote": "代码与模型权重许可分开，权重按 NVIDIA 模型条款；复现须固定 N1 对应版本，不能把新主分支当原论文实现。", "whyUseful": "复现应区分预训练两任务测试、后训练及视频增强分支，保留任务权重、部分计分、checkpoint选择和本体接口。论文提供模型/数据/基准入口，但发布资产不等于本轮已验证完整大规模预训练可复现。", "contribution": "GR00T N1将Eagle-2视觉语言模块与流匹配Diffusion Transformer相连，以本体专用编码/解码适配不同状态动作维度。混合真实遥操作、物理仿真、人类视频和生成视频，借潜在动作或逆动力学伪标签利用无动作视频，再进行任务后训练。", "limitations": "实机主要桌面操作，不证明行走或开放世界通用性。部分成功计分使百分比并非严格完整回合成功；仿真选择末五检查点最高分。预训练规模、数据混合与模型容量共同变化，不能全归于单一架构。", "caveats": "论文实机平台是 Fourier GR-1，不能替换成后来演示中的其他人形机器人。", "evidence": [ { "url": "https://arxiv.org/abs/2503.14734", "note": "2025-03-18 首发与模型摘要。" }, { "url": "https://arxiv.org/html/2503.14734v1", "note": "第 4 节分别给出 Panda/GR-1 仿真与 GR-1 实机评测。" }, { "url": "https://github.com/NVIDIA/Isaac-GR00T", "note": "官方可用代码与独立代码/权重许可；当前主分支为后续版本。" }, { "url": "https://huggingface.co/nvidia/GR00T-N1-2B", "note": "原 N1-2B 权重及 NVIDIA 模型许可入口。" }, { "url": "https://arxiv.org/html/2503.14734v1", "note": "§4.3：实机部分计分；仿真末五检查点取最高。" }, { "url": "https://arxiv.org/html/2503.14734v1", "note": "Table 3：76.8%为后训练真实任务平均；10%数据42.6%。" }, { "url": "https://arxiv.org/html/2503.14734v1", "note": "§4.4：预训练单独两任务，不能与后训练结果混为零样本。" }, { "url": "https://arxiv.org/html/2503.14734v1", "note": "PDF p.15, Tables 2–3 (visual inspection)：已核对仿真45.0%与实机76.8%属于不同表及设置；页面明确预训练测试允许0.5部分得分。" } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "人形机器人" ], "robotFilters": [ "Fourier GR-1" ], "verificationNote": "已核原论文、官方代码和原代模型卡；不采用模型卡中与论文不一致的后续架构描述。", "original": { "id": "arxiv-2503.14734", "originalSourceUrl": "https://arxiv.org/pdf/2503.14734v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2503.14734v2", "finalDownloadUrl": "https://arxiv.org/pdf/2503.14734", "sha256": "76e38cf679637b816b16edf7170340883dd09175bf334aeeaed49392777731b9", "bytes": 21021950, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:32.501619+00:00", "publicationDate": "2025/03/18", "publicationDateSourceUrl": "https://arxiv.org/abs/2503.14734", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2503.14734", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2503.14734", "pages": 36, "sourceIdentityStatus": "title_match", "sourceTitle": "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots", "sourceVersion": "2503.14734v2", "sourceVersionDate": "2025/03/27", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.840213+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2503.14734" }, "originalAnalysis": { "id": "arxiv-2503.14734", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090129+00:00", "sourceUrl": "https://arxiv.org/html/2503.14734v1", "sectionsRead": [ "§2–3", "§4.1–4.5", "Tables 2–3", "Appendix B, E" ], "methodsZh": "GR00T N1将Eagle-2视觉语言模块与流匹配Diffusion Transformer相连，以本体专用编码/解码适配不同状态动作维度。混合真实遥操作、物理仿真、人类视频和生成视频，借潜在动作或逆动力学伪标签利用无动作视频，再进行任务后训练。", "experimentsZh": "仿真覆盖RoboCasa24、DexMimicGen9及GR-1桌面24任务，每任务30/100/300示范；真实Fourier GR-1涵盖抓放、关节物体、工业和协作任务。多数实机任务10次，Pack Machinery五次并按物体完成数计分，使用阶段性部分得分。", "resultsZh": "作者表3全数据平均76.8%，Diffusion Policy46.4%；仅10%数据为42.6%，基线10.2%。仿真100示范设置平均45.0%，基线33.4%；这是后训练收益，不能当所有新任务零样本成功率。", "limitationsZh": "实机主要桌面操作，不证明行走或开放世界通用性。部分成功计分使百分比并非严格完整回合成功；仿真选择末五检查点最高分。预训练规模、数据混合与模型容量共同变化，不能全归于单一架构。", "reproductionZh": "复现应区分预训练两任务测试、后训练及视频增强分支，保留任务权重、部分计分、checkpoint选择和本体接口。论文提供模型/数据/基准入口，但发布资产不等于本轮已验证完整大规模预训练可复现。", "experimentType": "both", "robots": [ "Fourier GR-1" ], "evidenceNotes": [ { "section/page": "§4.3", "note": "实机部分计分；仿真末五检查点取最高。" }, { "section/page": "Table 3", "note": "76.8%为后训练真实任务平均；10%数据42.6%。" }, { "section/page": "§4.4", "note": "预训练单独两任务，不能与后训练结果混为零样本。" }, { "section/page": "PDF p.15, Tables 2–3 (visual inspection)", "note": "已核对仿真45.0%与实机76.8%属于不同表及设置；页面明确预训练测试允许0.5部分得分。" } ], "corrections": { "metricNote": "实机含阶段/物体部分得分，不应全部解释为完整回合成功率。" } }, "experimentNote": "仿真覆盖RoboCasa24、DexMimicGen9及GR-1桌面24任务，每任务30/100/300示范；真实Fourier GR-1涵盖抓放、关节物体、工业和协作任务。多数实机任务10次，Pack Machinery五次并按物体完成数计分，使用阶段性部分得分。", "analysisVerifiedAt": "2026-10-04T13:50:35.090129+00:00" }, { "id": "arxiv-2502.19645", "title": "Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success", "titleZh": "微调视觉—语言—动作模型：优化速度与成功率", "shortTitle": "OpenVLA-OFT", "date": "2025-02-27", "year": 2025, "url": "https://arxiv.org/abs/2502.19645", "paperUrl": "https://arxiv.org/abs/2502.19645", "projectUrl": "https://openvla-oft.github.io/", "codeUrl": "https://github.com/moojink/openvla-oft", "summary": "系统比较OpenVLA微调的自回归/并行解码、离散/连续动作及L1/扩散目标，形成并行预测动作块和连续L1回归的OFT。增加多视角、本体状态接口，并在双臂OFT+中用FiLM强化语言条件；无需另训练一个执行控制…", "abstractZh": "系统比较OpenVLA微调的自回归/并行解码、离散/连续动作及L1/扩散目标，形成并行预测动作块和连续L1回归的OFT。增加多视角、本体状态接口，并在双臂OFT+中用FiLM强化语言条件；无需另训练一个执行控制器。\nLIBERO四套任务每套500回合，区分过滤数据和不同输入配置。ALOHA四种真实双臂任务按任务独立微调，使用25步动作块；真实评测约10至24次/任务，按预先定义的部分完成分数及语言目标选择评分。", "category": "VLA / 高效微调", "tags": [ "OpenVLA-OFT", "视觉语言动作", "模仿学习", "操作与抓取" ], "directions": [ "视觉语言动作", "操作与抓取", "模仿学习" ], "tier": "recent", "experimentType": "both", "experimentNote": "LIBERO四套任务每套500回合，区分过滤数据和不同输入配置。ALOHA四种真实双臂任务按任务独立微调，使用25步动作块；真实评测约10至24次/任务，按预先定义的部分完成分数及语言目标选择评分。", "robots": [ "ALOHA" ], "robotFilters": [ "ALOHA" ], "robotNote": "摘要明确 ALOHA 双臂实机；LIBERO 仿真型号本轮未单独核实，未添加。", "codeStatus": "open", "status": "训练代码已开源", "trainingStatus": "训练代码已开源", "trainingNote": "官方 MIT 仓库提供 LIBERO/ALOHA 微调与评估说明、检查点；上游基础模型和数据条款应另查。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "MIT（代码；上游资源另行核对）", "contribution": "系统比较OpenVLA微调的自回归/并行解码、离散/连续动作及L1/扩散目标，形成并行预测动作块和连续L1回归的OFT。增加多视角、本体状态接口，并在双臂OFT+中用FiLM强化语言条件；无需另训练一个执行控制器。", "whyUseful": "应复用各表的数据/模态分组、动作块长度及近零动作过滤规则，同时测生成吞吐与重规划频率。真实任务需公开评分细则、语言选择与动作完成分开计数；未执行训练。", "limitations": "部分外部基准数引用原论文，训练过滤和输入模态不同。真实图中的完成百分比不是全部任务成功率。示范策略较一致，L1可能不保留多峰行为；FiLM的必要性在仿真与实机间不同，原因未明。", "caveats": "主要围绕 OpenVLA 和所选基准；更换任务、控制频率和数据预算需重新比较。", "evidence": [ { "url": "https://arxiv.org/abs/2502.19645", "note": "首发日期、方法与仿真/实机说明。" }, { "url": "https://github.com/moojink/openvla-oft", "note": "官方训练、评估、ALOHA 与 LIBERO 文档及 MIT。" }, { "url": "https://arxiv.org/pdf/2502.19645", "note": "Table I：95.3%与97.1%对应不同输入条件。" }, { "url": "https://arxiv.org/pdf/2502.19645", "note": "§VI; Table III; §VIII：部分完成评分、77.9Hz吞吐与0.321秒块延迟、多峰限制。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2025：把 VLA 微调策略作为独立研究对象，提升部署效率和任务成功率。", "freshness": "近年进展", "original": { "id": "arxiv-2502.19645", "originalSourceUrl": "https://arxiv.org/pdf/2502.19645v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2502.19645v2", "finalDownloadUrl": "https://arxiv.org/pdf/2502.19645", "sha256": "b860aa1206b6cfb0ce8be177f961379dd6a133d52cc74ac346636e0f4952a596", "bytes": 26994977, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:32.739501+00:00", "publicationDate": "2025/02/27", "publicationDateSourceUrl": "https://arxiv.org/abs/2502.19645", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2502.19645", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2502.19645", "pages": 24, "sourceIdentityStatus": "title_match", "sourceTitle": "Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success", "sourceVersion": "2502.19645v2", "sourceVersionDate": "2025/04/28", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.869880+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2502.19645" }, "originalAnalysis": { "id": "arxiv-2502.19645", "sourceUrl": "https://arxiv.org/pdf/2502.19645", "sectionsRead": [ "§III–V", "Table I", "§VI", "Table III", "§VIII" ], "methodsZh": "系统比较OpenVLA微调的自回归/并行解码、离散/连续动作及L1/扩散目标，形成并行预测动作块和连续L1回归的OFT。增加多视角、本体状态接口，并在双臂OFT+中用FiLM强化语言条件；无需另训练一个执行控制器。", "experimentsZh": "LIBERO四套任务每套500回合，区分过滤数据和不同输入配置。ALOHA四种真实双臂任务按任务独立微调，使用25步动作块；真实评测约10至24次/任务，按预先定义的部分完成分数及语言目标选择评分。", "resultsZh": "作者报告单第三人称输入OFT平均95.3%，含额外输入完整版97.1%，不能混作同配置增益。ALOHA上OFT+报告最高平均任务进度；A100平均100次请求吞吐77.9动作/秒、单块延迟0.321秒，原OpenVLA为1.8动作/秒。", "limitationsZh": "部分外部基准数引用原论文，训练过滤和输入模态不同。真实图中的完成百分比不是全部任务成功率。示范策略较一致，L1可能不保留多峰行为；FiLM的必要性在仿真与实机间不同，原因未明。", "reproductionZh": "应复用各表的数据/模态分组、动作块长度及近零动作过滤规则，同时测生成吞吐与重规划频率。真实任务需公开评分细则、语言选择与动作完成分开计数；未执行训练。", "experimentType": "both", "robots": [ "ALOHA" ], "evidenceNotes": [ { "section/page": "Table I", "note": "95.3%与97.1%对应不同输入条件。" }, { "section/page": "§VI; Table III; §VIII", "note": "部分完成评分、77.9Hz吞吐与0.321秒块延迟、多峰限制。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2502.01143", "title": "ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills", "titleZh": "ASAP：对齐仿真与真实物理以学习敏捷人形全身技能", "shortTitle": "ASAP", "date": "2025-02-03", "datePrecision": "day", "year": 2025, "url": "https://arxiv.org/abs/2502.01143", "paperUrl": "https://arxiv.org/abs/2502.01143", "codeUrl": "https://github.com/LeCAR-Lab/ASAP", "summary": "ASAP先把视频人体动作经SMPL重建、物理清洗及形态重定向变为机器人参考，以PPO学习跟踪；再用真实执行轨迹训练状态—动作条件的残差动作模型，使仿真复现真实响应。固定该残差模型改造训练动态后，重新微调原控制策…", "abstractZh": "ASAP先把视频人体动作经SMPL重建、物理清洗及形态重定向变为机器人参考，以PPO学习跟踪；再用真实执行轨迹训练状态—动作条件的残差动作模型，使仿真复现真实响应。固定该残差模型改造训练动态后，重新微调原控制策略。\n比较IsaacGym到IsaacSim/Genesis的迁移及真实Unitree G1；有系统辨识、直接残差状态动态等对照。真机因样本成本把23自由度残差缩成四个踝关节，用100段动作记录训练，并测试踢腿及未见LeBron动作。", "category": "人形机器人 / 实机适应", "tags": [ "残差动作模型", "Sim2Real", "动作跟踪", "敏捷运动" ], "directions": [ "人形机器人", "运动控制" ], "tier": "recent", "experimentType": "both", "robots": [ "Unitree G1" ], "robotNote": "", "codeStatus": "open", "status": "官方动作跟踪、残差动作训练与微调流程公开。", "trainingNote": "仓库分别提供运动跟踪、delta action model训练和策略微调命令；真实或跨仿真轨迹需包含对应动作。", "contribution": "ASAP先把视频人体动作经SMPL重建、物理清洗及形态重定向变为机器人参考，以PPO学习跟踪；再用真实执行轨迹训练状态—动作条件的残差动作模型，使仿真复现真实响应。固定该残差模型改造训练动态后，重新微调原控制策略。", "whyUseful": "先重现参考状态初始化和终止阈值课程，再在受控仿真验证残差是否恢复真实轨迹；独立检查训练/未见动作。须保留实际采样、控制频率和残差范围，真实部署先设关节热与负载限制。", "limitations": "需要动捕，完整23自由度残差需大量数据；真实结论主要基于四踝关节版本，不能视为全身动态均已辨识。采集曾导致两台G1不同程度损坏及电机过热，展示敏捷不等于可靠安全部署。", "evidence": [ { "url": "https://arxiv.org/abs/2502.01143", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://agile.human2humanoid.com/", "note": "官方四步流程、G1实机与跨仿真器实验说明。" }, { "url": "https://github.com/LeCAR-Lab/ASAP", "note": "README实际提供动作跟踪、残差模型训练及微调命令，不能只按旧TODO判定待开源。" }, { "url": "https://arxiv.org/pdf/2502.01143", "note": "III; IV-C：残差插入训练仿真、真实100段与四踝关节降维。" }, { "url": "https://arxiv.org/pdf/2502.01143", "note": "Table V; VIII：真实跟踪数字、动捕/样本及硬件损坏限制。" }, { "url": "https://arxiv.org/pdf/2502.01143v3", "note": "III; IV-C：残差插入训练仿真、真实100段与四踝关节降维。" }, { "url": "https://arxiv.org/pdf/2502.01143v3", "note": "Table V; VIII：真实跟踪数字、动捕/样本及硬件损坏限制。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2025：人形敏捷技能从纯随机化迁移走向真实数据驱动的物理对齐。", "projectUrl": null, "robotFilters": [ "Unitree G1" ], "original": { "id": "arxiv-2502.01143", "originalSourceUrl": "https://arxiv.org/pdf/2502.01143v3", "versionedOriginalUrl": "https://arxiv.org/pdf/2502.01143v3", "finalDownloadUrl": "https://arxiv.org/pdf/2502.01143", "sha256": "8ca60229d7e560769ceece6cb5571d8e6f528c6c485f631b7fd5162d503b7db0", "bytes": 36214501, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:38.854057+00:00", "publicationDate": "2025/02/03", "publicationDateSourceUrl": "https://arxiv.org/abs/2502.01143", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2502.01143", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2502.01143", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills", "sourceVersion": "2502.01143v3", "sourceVersionDate": "2025/04/26", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.907377+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2502.01143" }, "originalAnalysis": { "id": "arxiv-2502.01143", "sourceUrl": "https://arxiv.org/pdf/2502.01143v3", "sectionsRead": [ "II", "III-A–C", "IV-A–C", "V", "VIII", "Tables IV–V" ], "methodsZh": "ASAP先把视频人体动作经SMPL重建、物理清洗及形态重定向变为机器人参考，以PPO学习跟踪；再用真实执行轨迹训练状态—动作条件的残差动作模型，使仿真复现真实响应。固定该残差模型改造训练动态后，重新微调原控制策略。", "experimentsZh": "比较IsaacGym到IsaacSim/Genesis的迁移及真实Unitree G1；有系统辨识、直接残差状态动态等对照。真机因样本成本把23自由度残差缩成四个踝关节，用100段动作记录训练，并测试踢腿及未见LeBron动作。", "resultsZh": "真实踢腿全局关节点误差从61.2降至50.2，未见LeBron从159降至112；速度和加速度误差也下降。跨模拟器复杂动作成功与跟踪改善，说明校正动作通道可成为动态差异的实用代理，而不是只改模型物理参数。", "limitationsZh": "需要动捕，完整23自由度残差需大量数据；真实结论主要基于四踝关节版本，不能视为全身动态均已辨识。采集曾导致两台G1不同程度损坏及电机过热，展示敏捷不等于可靠安全部署。", "reproductionZh": "先重现参考状态初始化和终止阈值课程，再在受控仿真验证残差是否恢复真实轨迹；独立检查训练/未见动作。须保留实际采样、控制频率和残差范围，真实部署先设关节热与负载限制。", "experimentType": "both", "robots": [ "Unitree G1" ], "corrections": [ "实机采用四自由度踝关节残差模型，不能误写为完整23自由度在线适配。" ], "evidenceNotes": [ { "note": "残差插入训练仿真、真实100段与四踝关节降维。", "section": "III; IV-C" }, { "note": "真实跟踪数字、动捕/样本及硬件损坏限制。", "section": "Table V; VIII" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2502.01143v3", "originalSha256": "8ca60229d7e560769ceece6cb5571d8e6f528c6c485f631b7fd5162d503b7db0" }, "experimentNote": "比较IsaacGym到IsaacSim/Genesis的迁移及真实Unitree G1；有系统辨识、直接残差状态动态等对照。真机因样本成本把23自由度残差缩成四个踝关节，用100段动作记录训练，并测试踢腿及未见LeBron动作。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2501.09747", "title": "FAST: Efficient Action Tokenization for Vision-Language-Action Models", "titleZh": "FAST：视觉—语言—动作模型的高效动作词元化", "shortTitle": "FAST", "date": "2025-01-16", "year": 2025, "url": "https://arxiv.org/abs/2501.09747", "paperUrl": "https://arxiv.org/abs/2501.09747", "projectUrl": null, "codeUrl": "https://huggingface.co/physical-intelligence/fast", "summary": "FAST先将一秒动作段按维度做离散余弦变换，量化频域系数，按低频优先展开，再用BPE压缩为自回归动作词元。FAST+把词典训练在约百万跨形态动作段上；无需另训神经压缩网络，但量化仍有损失，BPE本身才是无损编码…", "abstractZh": "FAST先将一秒动作段按维度做离散余弦变换，量化频域系数，按低频优先展开，再用BPE压缩为自回归动作词元。FAST+把词典训练在约百万跨形态动作段上；无需另训神经压缩网络，但量化仍有损失，BPE本身才是无损编码。\n以π0和OpenVLA骨干对比逐维分箱、FSQ及扩散控制，涵盖LIBERO和六类实机评估。硬件包括UR5e、ARX双臂、Trossen ViperX；DROID在新场景44次定量测试，另三校园演示仅定性。大型通用策略训练约一万小时动作数据。", "category": "VLA / 动作表示", "tags": [ "FAST", "视觉语言动作", "模仿学习", "操作与抓取" ], "directions": [ "视觉语言动作", "操作与抓取", "模仿学习" ], "tier": "recent", "experimentType": "both", "experimentNote": "以π0和OpenVLA骨干对比逐维分箱、FSQ及扩散控制，涵盖LIBERO和六类实机评估。硬件包括UR5e、ARX双臂、Trossen ViperX；DROID在新场景44次定量测试，另三校园演示仅定性。大型通用策略训练约一万小时动作数据。", "robots": [ "UR5e", "ARX双臂（型号未细化）", "Trossen ViperX双臂", "Franka（DROID评估）" ], "robotFilters": [ "UR5e", "Franka（型号未注明）" ], "robotNote": "原文另有双臂 ARX，但具体型号未核实，未塞入型号筛选；Franka 评估平台的更细型号未核实。", "codeStatus": "open", "status": "训练代码已开源", "trainingStatus": "训练代码已开源", "trainingNote": "官方 Hugging Face 发布 FAST+ 和自定义词元器训练代码，标注 Apache-2.0；此范围不代表全部 VLA 预训练数据公开。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "Apache-2.0（FAST 词元器）", "contribution": "FAST先将一秒动作段按维度做离散余弦变换，量化频域系数，按低频优先展开，再用BPE压缩为自回归动作词元。FAST+把词典训练在约百万跨形态动作段上；无需另训神经压缩网络，但量化仍有损失，BPE本身才是无损编码。", "whyUseful": "复现需动作分位归一化、DCT量化尺度、低频展开及BPE词典；常用尺度10、词表1024。DROID设置筛选75K成功轨迹、去掉零动作，15步预测并执行8或15步；发布tokenizer不等于全部私有训练数据开放。", "limitations": "实机政策验证集中静态机械臂，手/人形/移动平台主要是离线压缩测试。FAST+词典混合包含评测任务的数据，不能称完全任务外训练；多个评价使用物体比例或任务进度而非整回合成功，衣物任务仍有高质量数据微调。", "caveats": "压缩存在精度与词元长度取舍；高频任务仍需可靠控制栈和合适训练数据。", "evidence": [ { "url": "https://arxiv.org/abs/2501.09747", "note": "首发日期与标题。" }, { "url": "https://arxiv.org/html/2501.09747v1", "note": "§VI、附录E：UR5e、Trossen ViperX、ARX 与 DROID 实机任务。" }, { "url": "https://huggingface.co/physical-intelligence/fast", "note": "官方词元器、训练函数和 Apache-2.0 标签。" }, { "url": "https://arxiv.org/html/2501.09747v1", "note": "V-B; Table I：DCT量化有损；折衣700→53词元。" }, { "url": "https://arxiv.org/html/2501.09747v1", "note": "VI-E–F：五倍为训练GPU小时；推理FAST约750ms、扩散约100ms。" }, { "url": "https://arxiv.org/html/2501.09747v1", "note": "Appendix E：DROID 44次定量；三校园仅定性；任务评分口径不同。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2025：以时序压缩缓解自回归 VLA 在高频动作序列上的表示瓶颈。", "freshness": "近年进展", "original": { "id": "arxiv-2501.09747", "originalSourceUrl": "https://arxiv.org/pdf/2501.09747v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2501.09747v1", "finalDownloadUrl": "https://arxiv.org/pdf/2501.09747", "sha256": "3739b31f5fecdde371509ff5bb13619979734e894a255a9b264253f4cc53934a", "bytes": 7468382, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:43.849473+00:00", "publicationDate": "2025/01/16", "publicationDateSourceUrl": "https://arxiv.org/abs/2501.09747", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2501.09747", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2501.09747", "pages": 19, "sourceIdentityStatus": "title_match", "sourceTitle": "FAST: Efficient Action Tokenization for Vision-Language-Action Models", "sourceVersion": "2501.09747v1", "sourceVersionDate": "2025/01/16", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.918876+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2501.09747" }, "originalAnalysis": { "id": "arxiv-2501.09747", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458248+00:00", "sourceUrl": "https://arxiv.org/html/2501.09747v1", "sectionsRead": [ "V-B–C", "VI-A–F", "Table I", "VII Discussion", "Appendix C–E" ], "methodsZh": "FAST先将一秒动作段按维度做离散余弦变换，量化频域系数，按低频优先展开，再用BPE压缩为自回归动作词元。FAST+把词典训练在约百万跨形态动作段上；无需另训神经压缩网络，但量化仍有损失，BPE本身才是无损编码。", "experimentsZh": "以π0和OpenVLA骨干对比逐维分箱、FSQ及扩散控制，涵盖LIBERO和六类实机评估。硬件包括UR5e、ARX双臂、Trossen ViperX；DROID在新场景44次定量测试，另三校园演示仅定性。大型通用策略训练约一万小时动作数据。", "resultsZh": "作者报告50Hz双臂折衣的一秒动作从700词元压到53，压缩13.2倍；通用π0-FAST达到类似扩散π0表现所需GPU时约少五倍。但4090推理约750毫秒/段，对扩散约100毫秒，训练快不能写成运行快。", "limitationsZh": "实机政策验证集中静态机械臂，手/人形/移动平台主要是离线压缩测试。FAST+词典混合包含评测任务的数据，不能称完全任务外训练；多个评价使用物体比例或任务进度而非整回合成功，衣物任务仍有高质量数据微调。", "reproductionZh": "复现需动作分位归一化、DCT量化尺度、低频展开及BPE词典；常用尺度10、词表1024。DROID设置筛选75K成功轨迹、去掉零动作，15步预测并执行8或15步；发布tokenizer不等于全部私有训练数据开放。", "experimentType": "both", "robots": [ "UR5e", "ARX双臂（型号未细化）", "Trossen ViperX双臂", "Franka（DROID评估）" ], "evidenceNotes": [ { "section": "V-B; Table I", "note": "DCT量化有损；折衣700→53词元。" }, { "section": "VI-E–F", "note": "五倍为训练GPU小时；推理FAST约750ms、扩散约100ms。" }, { "section": "Appendix E", "note": "DROID 44次定量；三校园仅定性；任务评分口径不同。" } ], "corrections": [ "速度贡献主要训练效率，不能宣称FAST推理比扩散更快。" ], "sourceVersion": "2501.09747v1", "originalSha256": "3739b31f5fecdde371509ff5bb13619979734e894a255a9b264253f4cc53934a" }, "analysisVerifiedAt": "2026-10-04T13:45:34.458248+00:00" }, { "id": "arxiv-2412.04453", "title": "NaVILA: Legged Robot Vision-Language-Action Model for Navigation", "titleZh": "NaVILA：面向足式机器人导航的视觉语言动作模型", "date": "2024-12-05", "datePrecision": "day", "year": 2024, "url": "https://arxiv.org/abs/2412.04453", "paperUrl": "https://arxiv.org/abs/2412.04453", "codeUrl": "https://github.com/AnjieCheng/NaVILA", "summary": "高层VILA以当前图像和历史帧理解语言，输出带距离/角度的自然语言移动指令，解析后交给独立的视觉强化学习步态控制器。训练混合仿真导航、人类游览视频、辅助导航和VQA；高层语义与低层避障的接口便于跨本体组合。", "abstractZh": "高层VILA以当前图像和历史帧理解语言，输出带距离/角度的自然语言移动指令，解析后交给独立的视觉强化学习步态控制器。训练混合仿真导航、人类游览视频、辅助导航和VQA；高层语义与低层避障的接口便于跨本体组合。\n评估R2R/RxR未见场景和ScanQA，并把1077条可通行R2R路线导入Isaac物理仿真，比较Go2/H1的感知和盲控。实机25条指令各重复三次，涵盖工作区、家与室外；另有G1展示。", "category": "具身导航 / 足式 VLA", "tags": [ "NaVILA", "视觉语言动作", "足式导航", "分层控制" ], "directions": [ "视觉语言动作", "运动控制", "导航与建图" ], "tier": "recent", "experimentType": "both", "robots": [ "Unitree Go2", "Unitree G1" ], "robotFilters": [ "Unitree Go2", "Unitree G1" ], "codeStatus": "open", "status": "实现已开源", "trainingNote": "高层官方仓库已提供训练脚本、评测和权重入口；低层训练另在 legged-loco 仓库。人类视频仅发布 ID/标注，原始视频受版权限制，复现需自行取得。", "contribution": "高层VILA以当前图像和历史帧理解语言，输出带距离/角度的自然语言移动指令，解析后交给独立的视觉强化学习步态控制器。训练混合仿真导航、人类游览视频、辅助导航和VQA；高层语义与低层避障的接口便于跨本体组合。", "whyUseful": "复现要分开经典VLN的理想执行、Isaac接触控制和真实试验，保留路线筛选与成功判定。实机使用8帧记忆，64帧的离线问答收益不能直接套用；量化延迟在RTX4090测量，本轮未复现控制栈。", "limitations": "实机指令数量有限，存在目标位置误判，历史帧受计算限制。部署把Go2图像传到服务器，动作间等待约一秒；直接机载VLA为未来方向。H1出现在仿真比较，不能和G1实机展示混写。", "license": "Apache-2.0（高层 NaVILA）；MIT（低层 legged-loco）；模型/数据另核", "evidence": [ { "url": "https://arxiv.org/abs/2412.04453", "note": "首版 2024-12-05，2025 修订；核对两级方法与实验范围。" }, { "url": "https://arxiv.org/html/2412.04453v2", "note": "§III-C/D、表 IV/VI 区分 Go2/H1 仿真与 Go2/Booster T1 实机。" }, { "url": "https://navila-bot.github.io/", "note": "作者项目页直链高层/低层代码与基准。" }, { "url": "https://github.com/AnjieCheng/NaVILA", "note": "高层 Apache-2.0；训练脚本、数据标注与视频版权限制有明确说明。" }, { "url": "https://github.com/yang-zj1026/legged-loco", "note": "低层训练 MIT 仓库，说明 Isaac Lab 中 Go2/H1 训练。" }, { "url": "https://arxiv.org/html/2412.04453v1", "note": "§3.2 / Table 4：Go2/H1是物理仿真；低层感知提升成功率。" }, { "url": "https://arxiv.org/html/2412.04453v1", "note": "§3.3：25指令×3；正文另提G1实机，无需重训高层。" }, { "url": "https://arxiv.org/html/2412.04453v1", "note": "Appendix I：图像传服务器、约1秒间隔；机载VLA尚为未来工作。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "以中层语言动作连接 VLA 与足式运动技能，推动语言导航走向物理机器人执行。", "experimentNote": "评估R2R/RxR未见场景和ScanQA，并把1077条可通行R2R路线导入Isaac物理仿真，比较Go2/H1的感知和盲控。实机25条指令各重复三次，涵盖工作区、家与室外；另有G1展示。", "robotNote": "Unitree H1 为仿真；Unitree Go2 兼有仿真/实机；Booster T1 为实机。", "projectUrl": null, "shortTitle": "", "original": { "id": "arxiv-2412.04453", "originalSourceUrl": "https://arxiv.org/pdf/2412.04453v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2412.04453v2", "finalDownloadUrl": "https://arxiv.org/pdf/2412.04453", "sha256": "e02fcccae8aad56cf33234882de06b97256b76aad86a011171b83db40790bc8c", "bytes": 9981598, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:43.906816+00:00", "publicationDate": "2024/12/05", "publicationDateSourceUrl": "https://arxiv.org/abs/2412.04453", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2412.04453", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2412.04453", "pages": 17, "sourceIdentityStatus": "title_match", "sourceTitle": "NaVILA: Legged Robot Vision-Language-Action Model for Navigation", "sourceVersion": "2412.04453v2", "sourceVersionDate": "2025/02/17", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.929360+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2412.04453" }, "originalAnalysis": { "id": "arxiv-2412.04453", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090131+00:00", "sourceUrl": "https://arxiv.org/html/2412.04453v1", "sectionsRead": [ "§2", "§3.1–3.3", "Tables 4–5", "Appendices I–J" ], "methodsZh": "高层VILA以当前图像和历史帧理解语言，输出带距离/角度的自然语言移动指令，解析后交给独立的视觉强化学习步态控制器。训练混合仿真导航、人类游览视频、辅助导航和VQA；高层语义与低层避障的接口便于跨本体组合。", "experimentsZh": "评估R2R/RxR未见场景和ScanQA，并把1077条可通行R2R路线导入Isaac物理仿真，比较Go2/H1的感知和盲控。实机25条指令各重复三次，涵盖工作区、家与室外；另有G1展示。", "resultsZh": "作者报告Go2物理仿真成功率由盲控36.2%升至50.2%，H1由24.4%升至45.3%；真实指令测试汇总88%。说明低层环境感知明显影响结果，不能只凭高层语言模型得分判断机器人可执行性。", "limitationsZh": "实机指令数量有限，存在目标位置误判，历史帧受计算限制。部署把Go2图像传到服务器，动作间等待约一秒；直接机载VLA为未来方向。H1出现在仿真比较，不能和G1实机展示混写。", "reproductionZh": "复现要分开经典VLN的理想执行、Isaac接触控制和真实试验，保留路线筛选与成功判定。实机使用8帧记忆，64帧的离线问答收益不能直接套用；量化延迟在RTX4090测量，本轮未复现控制栈。", "experimentType": "both", "robots": [ "Unitree Go2", "Unitree G1" ], "evidenceNotes": [ { "section/page": "§3.2 / Table 4", "note": "Go2/H1是物理仿真；低层感知提升成功率。" }, { "section/page": "§3.3", "note": "25指令×3；正文另提G1实机，无需重训高层。" }, { "section/page": "Appendix I", "note": "图像传服务器、约1秒间隔；机载VLA尚为未来工作。" } ], "corrections": { "robotNote": "Go2、G1有实机证据；H1为本论文仿真评估。" } }, "analysisVerifiedAt": "2026-10-04T13:50:35.090131+00:00" }, { "id": "arxiv-2410.24164", "title": "π0: A Vision-Language-Action Flow Model for General Robot Control", "titleZh": "π0：面向通用机器人控制的视觉－语言－动作流模型", "shortTitle": "π0", "date": "2024-10-31", "year": 2024, "url": "https://arxiv.org/abs/2410.24164", "codeUrl": "https://github.com/Physical-Intelligence/openpi", "category": "视觉语言动作", "tags": [ "流匹配", "动作分块", "跨本体", "灵巧操作" ], "tier": "foundation", "summary": "基于PaliGemma视觉语言主干加入较小动作专家，以条件流匹配生成50步连续动作块，多相机、指令和关节状态共同输入。先在跨机器人广泛数据上预训练，再用质量与策略一致性更高的任务数据后训练，动作维度填充兼容多个…", "abstractZh": "基于PaliGemma视觉语言主干加入较小动作专家，以条件流匹配生成50步连续动作块，多相机、指令和关节状态共同输入。先在跨机器人广泛数据上预训练，再用质量与策略一致性更高的任务数据后训练，动作维度填充兼容多个平台。\n自有约一万小时、七种配置、68任务数据与OXE等混合；评测预训练内直接指令任务、语言遵循、1/5/10小时微调及复杂长程任务。多类实机包括UR5e、Franka、双Trossen/ARX和移动平台，各主评测通常十次。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "real", "robots": [ "UR5e", "Franka（正文未在此处细分型号）", "Trossen双臂", "ARX双臂", "Mobile Trossen", "Mobile Fibocom" ], "codeStatus": "open", "status": "官方 openpi 提供模型、检查点与训练/推理实现。", "trainingNote": "代码 Apache-2.0，另含 Gemma 条款；论文大规模自采数据未全部公开。开放版本与原文内部系统并非完全相同。", "whyUseful": "复现要同时控制预训练混合、平台动作定义、后训练示范品质与评分规程，不能只复制流匹配结构。应逐项注明是否预训练见过、是否有高层提示，结果均为作者报告。", "contribution": "基于PaliGemma视觉语言主干加入较小动作专家，以条件流匹配生成50步连续动作块，多相机、指令和关节状态共同输入。先在跨机器人广泛数据上预训练，再用质量与策略一致性更高的任务数据后训练，动作维度填充兼容多个平台。", "limitations": "基础模型的直接评测主要为预训练含有的任务；长任务部分需要高层语言指导，不能全部称自主发现流程。全模型和对照训练预算不一，虽补有计算量匹配控制；十次试验不足以证明长期可靠性。", "caveats": "实机配置按原文 V-C；平台类别与具体硬件实例数量不能混用。", "evidence": [ { "url": "https://arxiv.org/html/2410.24164v1", "note": "首发日期；第 IV–VI 节架构、数据、硬件与实机实验。" }, { "url": "https://github.com/Physical-Intelligence/openpi", "note": "Apache-2.0、Gemma 条款及公开模型范围。" }, { "url": "https://arxiv.org/pdf/2410.24164", "note": "§IV–V; Figure 5：50步流匹配、七配置与机器人列表。" }, { "url": "https://arxiv.org/pdf/2410.24164", "note": "§VI-A,D; Figure 13：每任务十次、部分完成计分和预训练任务覆盖。" } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "视觉语言动作", "操作与抓取", "灵巧手" ], "robotFilters": [ "UR5e", "Franka（型号未注明）", "Mobile Trossen", "Mobile Fibocom" ], "verificationNote": "已核原文机器人系统细节和官方实现；未把开放权重等同于完整开放数据。", "original": { "id": "arxiv-2410.24164", "originalSourceUrl": "https://arxiv.org/pdf/2410.24164v4", "versionedOriginalUrl": "https://arxiv.org/pdf/2410.24164v4", "finalDownloadUrl": "https://arxiv.org/pdf/2410.24164", "sha256": "fbdfba56258bdbd220207ba63410c6c943e43a73e97a7a5a47ca0bc74204f821", "bytes": 7893845, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:48.490593+00:00", "publicationDate": "2024/10/31", "publicationDateSourceUrl": "https://arxiv.org/abs/2410.24164", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2410.24164", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2410.24164", "pages": 17, "sourceIdentityStatus": "title_match", "sourceTitle": "$\\pi_0$: A Vision-Language-Action Flow Model for General Robot Control", "sourceVersion": "2410.24164v4", "sourceVersionDate": "2026/01/08", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.939304+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2410.24164" }, "originalAnalysis": { "id": "arxiv-2410.24164", "sourceUrl": "https://arxiv.org/pdf/2410.24164", "sectionsRead": [ "§IV–V", "§VI-A–D", "Figures 5,11–13", "§VII" ], "methodsZh": "基于PaliGemma视觉语言主干加入较小动作专家，以条件流匹配生成50步连续动作块，多相机、指令和关节状态共同输入。先在跨机器人广泛数据上预训练，再用质量与策略一致性更高的任务数据后训练，动作维度填充兼容多个平台。", "experimentsZh": "自有约一万小时、七种配置、68任务数据与OXE等混合；评测预训练内直接指令任务、语言遵循、1/5/10小时微调及复杂长程任务。多类实机包括UR5e、Franka、双Trossen/ARX和移动平台，各主评测通常十次。", "resultsZh": "作者报告基础模型在五项直接评测中明显优于所测对照；等计算量版本仍领先。后训练后，叠衣、清桌、装盒等复杂任务平均进度均超过最大分数一半。分数可给部分完成，不能改写成所有复杂任务成功率超过50%。", "limitationsZh": "基础模型的直接评测主要为预训练含有的任务；长任务部分需要高层语言指导，不能全部称自主发现流程。全模型和对照训练预算不一，虽补有计算量匹配控制；十次试验不足以证明长期可靠性。", "reproductionZh": "复现要同时控制预训练混合、平台动作定义、后训练示范品质与评分规程，不能只复制流匹配结构。应逐项注明是否预训练见过、是否有高层提示，结果均为作者报告。", "experimentType": "real", "robots": [ "UR5e", "Franka（正文未在此处细分型号）", "Trossen双臂", "ARX双臂", "Mobile Trossen", "Mobile Fibocom" ], "evidenceNotes": [ { "section/page": "§IV–V; Figure 5", "note": "50步流匹配、七配置与机器人列表。" }, { "section/page": "§VI-A,D; Figure 13", "note": "每任务十次、部分完成计分和预训练任务覆盖。" } ], "corrections": [ { "field": "resultsNote", "value": "复杂任务图13报告平均进度分数而非二元成功率", "reason": "原文明确分数允许部分完成。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04" }, "experimentNote": "自有约一万小时、七种配置、68任务数据与OXE等混合；评测预训练内直接指令任务、语言遵循、1/5/10小时微调及复杂长程任务。多类实机包括UR5e、Franka、双Trossen/ARX和移动平台，各主评测通常十次。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2410.07864", "title": "RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation", "titleZh": "RDT-1B：面向双臂操作的扩散基础模型", "shortTitle": "RDT-1B", "date": "2024-10-10", "year": 2024, "url": "https://arxiv.org/abs/2410.07864", "paperUrl": "https://arxiv.org/abs/2410.07864", "projectUrl": "https://rdt-robotics.github.io/", "codeUrl": "https://github.com/thu-ml/RoboticsDiffusionTransformer", "summary": "RDT把动作建模为多模态连续分布，用扩散Transformer预测动作块；物理状态/动作使用傅里叶特征，语言与图像用预训练编码。跨机器人数据按物理含义填入统一动作槽并掩码，减少异构维度混淆，再在目标双臂平台微调…", "abstractZh": "RDT把动作建模为多模态连续分布，用扩散Transformer预测动作块；物理状态/动作使用傅里叶特征，语言与图像用预训练编码。跨机器人数据按物理含义填入统一动作槽并掩码，减少异构维度混淆，再在目标双臂平台微调。\n46数据集、百万余轨迹预训练，目标集6000多轨迹、300多任务；重点实机评测未见杯子和房间、指定手与水量、五示范交接、一示范叠衣及操纵遥控器。硬件为AgileX制造的Cobot Mobile ALOHA。", "category": "VLA / 双臂扩散策略", "tags": [ "RDT-1B", "视觉语言动作", "模仿学习", "操作与抓取" ], "directions": [ "视觉语言动作", "操作与抓取", "模仿学习" ], "tier": "recent", "experimentType": "real", "experimentNote": "46数据集、百万余轨迹预训练，目标集6000多轨迹、300多任务；重点实机评测未见杯子和房间、指定手与水量、五示范交接、一示范叠衣及操纵遥控器。硬件为AgileX制造的Cobot Mobile ALOHA。", "robots": [ "Cobot Mobile ALOHA（AgileX）" ], "robotFilters": [ "Mobile ALOHA" ], "robotNote": "AgileX 制造的 Cobot Mobile ALOHA；底盘仅便于搬运，实验是静态双臂操作。", "codeStatus": "open", "status": "训练代码已开源", "trainingStatus": "训练代码已开源", "trainingNote": "官方训练/微调实现与模型入口公开；代码 MIT，预训练数据各来源的权限仍应分别核对。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "MIT（代码；数据依各来源）", "contribution": "RDT把动作建模为多模态连续分布，用扩散Transformer预测动作块；物理状态/动作使用傅里叶特征，语言与图像用预训练编码。跨机器人数据按物理含义填入统一动作槽并掩码，减少异构维度混淆，再在目标双臂平台微调。", "whyUseful": "优先在公开目标数据做小模型或微调，严格复核统一动作各槽的单位、mask和相机历史；分别报告总成功、正确手、正确水量。基线双臂适配、动作块及目标集划分必须匹配。", "limitations": "移动底盘仅用于搬运场景，训练和测试都是静态双臂操作；无法据名称声称移动操作能力。不同任务试验数有限，综合表现还平均附加指令指标；预训练48张H100一月，完整复现成本高。", "caveats": "主实验聚焦静态双臂；大型模型和多源数据带来复现成本，不能由硬件名称推断移动能力。", "evidence": [ { "url": "https://arxiv.org/abs/2410.07864", "note": "首发日期与方法概述。" }, { "url": "https://arxiv.org/html/2410.07864v2", "note": "附录 G：AgileX Cobot Mobile ALOHA；明确未使用自主移动。" }, { "url": "https://github.com/thu-ml/RoboticsDiffusionTransformer", "note": "官方代码和训练说明；LICENSE 为 MIT。" }, { "url": "https://arxiv.org/pdf/2410.07864", "note": "4; 5; Table 3：统一物理动作空间、预训练规模和实机成绩。" }, { "url": "https://arxiv.org/pdf/2410.07864", "note": "Appendix G–H：精确硬件、静态双臂范围及综合评分。" }, { "url": "https://arxiv.org/pdf/2410.07864v2", "note": "4; 5; Table 3：统一物理动作空间、预训练规模和实机成绩。" }, { "url": "https://arxiv.org/pdf/2410.07864v2", "note": "Appendix G–H：精确硬件、静态双臂范围及综合评分。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2024：将大规模扩散 Transformer 与统一动作空间引入双臂通用策略。", "freshness": "近年进展", "original": { "id": "arxiv-2410.07864", "originalSourceUrl": "https://arxiv.org/pdf/2410.07864v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2410.07864v2", "finalDownloadUrl": "https://arxiv.org/pdf/2410.07864", "sha256": "fabfa885b63f2d68b8e7868549563c346d4323a596d24ee3e43d0a62b4ef9667", "bytes": 9699239, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:36:55.960932+00:00", "publicationDate": "2024/10/10", "publicationDateSourceUrl": "https://arxiv.org/abs/2410.07864", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2410.07864", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2410.07864", "pages": 28, "sourceIdentityStatus": "title_match", "sourceTitle": "RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation", "sourceVersion": "2410.07864v2", "sourceVersionDate": "2025/03/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.948062+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2410.07864" }, "originalAnalysis": { "id": "arxiv-2410.07864", "sourceUrl": "https://arxiv.org/pdf/2410.07864v2", "sectionsRead": [ "4.1–4.2", "5", "Appendix G–H", "Tables 1–3" ], "methodsZh": "RDT把动作建模为多模态连续分布，用扩散Transformer预测动作块；物理状态/动作使用傅里叶特征，语言与图像用预训练编码。跨机器人数据按物理含义填入统一动作槽并掩码，减少异构维度混淆，再在目标双臂平台微调。", "experimentsZh": "46数据集、百万余轨迹预训练，目标集6000多轨迹、300多任务；重点实机评测未见杯子和房间、指定手与水量、五示范交接、一示范叠衣及操纵遥控器。硬件为AgileX制造的Cobot Mobile ALOHA。", "resultsZh": "叠短裤一示范后成功68%，从头RDT40%；五示范交接40%，从头16%；未见房间倒水总成功62.5/100/62.5。DPM-Solver++把100步降至5步，4090约每秒6个动作块，381是动作采样吞吐而非381Hz闭环视觉更新。", "limitationsZh": "移动底盘仅用于搬运场景，训练和测试都是静态双臂操作；无法据名称声称移动操作能力。不同任务试验数有限，综合表现还平均附加指令指标；预训练48张H100一月，完整复现成本高。", "reproductionZh": "优先在公开目标数据做小模型或微调，严格复核统一动作各槽的单位、mask和相机历史；分别报告总成功、正确手、正确水量。基线双臂适配、动作块及目标集划分必须匹配。", "experimentType": "real", "robots": [ "Cobot Mobile ALOHA（AgileX）" ], "corrections": [ "虽然平台名含Mobile，论文明确没有训练或推理自主移动功能。", "381Hz是动作吞吐，动作块约6Hz，非视觉闭环控制频率。" ], "evidenceNotes": [ { "note": "统一物理动作空间、预训练规模和实机成绩。", "section": "4; 5; Table 3" }, { "note": "精确硬件、静态双臂范围及综合评分。", "section": "Appendix G–H" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2410.07864v2", "originalSha256": "fabfa885b63f2d68b8e7868549563c346d4323a596d24ee3e43d0a62b4ef9667" }, "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2409.13678", "title": "SoloParkour: Constrained Reinforcement Learning for Visual Locomotion from Privileged Experience", "date": "2024-09-20", "url": "https://arxiv.org/abs/2409.13678v1", "titleZh": "SoloParkour：用特权经验学习受约束的视觉跑酷控制", "abstractZh": "先学带特权地形信息的受约束策略，再用其经验初始化离策略视觉强化学习；让视觉策略适应有限观测，同时保留关节和姿态约束。\nIsaac Gym训练后部署Solo-12，Raspberry Pi 5上50Hz策略、D405深度相机；比较DAgger/BC蒸馏并统计地形完成与约束违反。", "summary": "先学带特权地形信息的受约束策略，再用其经验初始化离策略视觉强化学习；让视觉策略适应有限观测，同时保留关节和姿态约束。", "experimentType": "both", "robots": [ "Solo-12" ], "tags": [ "运动控制", "强化学习", "安全与评估" ], "limitations": "作者报告实机可能触及未建模的电池电流上限，并怀疑试验中损伤电池；满足电机限制不足以保证供电安全，训练地形仍需手工设计。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的受约束视觉运动研究，特别适合分析策略约束与电池限制的差别。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2409.13678v1", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2409.13678v1", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/Gepetto/SoloParkour", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/Gepetto/SoloParkour", "category": "运动控制 / 强化学习", "year": 2024, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "三阶段 train.py 命令已发布。 提供经验生成脚本；未核实单独公开训练数据包。 README 使用用户训练得到的 checkpoint；未核实现成权重下载。 三阶段命令依赖前一阶段权重和回放数据；源码可用不代表已提供完整实机启动包。 本次未运行训练、复现实验或实机控制。", "directions": [ "运动控制", "强化学习", "安全与评估" ], "robotFilters": [ "Solo-12" ], "original": { "id": "arxiv-2409.13678", "originalSourceUrl": "https://arxiv.org/pdf/2409.13678v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2409.13678v1", "finalDownloadUrl": "https://arxiv.org/pdf/2409.13678v1", "sha256": "f903a60fc167acc21a4a99dc010bbf51634561ac6a278375001e5548cd0c17bd", "bytes": 3718278, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:31.106836+00:00", "publicationDate": "2024/09/20", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2409.13678v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2409.13678v1", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "SoloParkour: Constrained Reinforcement Learning for Visual Locomotion from Privileged Experience", "sourceVersion": "2409.13678v1", "sourceVersionDate": "2024/09/20", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2409.13678v1", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2409.13678" }, "originalAnalysis": { "id": "arxiv-2409.13678", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2409.13678v1", "sourceVersion": "2409.13678v1", "sectionsRead": [ "PDF第4页选段（提取文本L188–201）", "PDF第4–5页选段（提取文本L202–219）", "PDF第5页选段（提取文本L251–268）", "PDF第6–7页选段（提取文本L332–345）", "PDF第8页选段（提取文本L393–410）", "PDF第6页选段（提取文本L294–329）", "PDF第7–8页选段（提取文本L362–391）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "先学带特权地形信息的受约束策略，再用其经验初始化离策略视觉强化学习；让视觉策略适应有限观测，同时保留关节和姿态约束。", "experimentsZh": "Isaac Gym训练后部署Solo-12，Raspberry Pi 5上50Hz策略、D405深度相机；比较DAgger/BC蒸馏并统计地形完成与约束违反。", "resultsZh": "实机可爬40厘米台阶、跨35厘米缝隙并从20厘米障碍下穿过；无约束视觉RL虽可能走得更远，却更频繁违反硬件约束。", "limitationsZh": "作者报告实机可能触及未建模的电池电流上限，并怀疑试验中损伤电池；满足电机限制不足以保证供电安全，训练地形仍需手工设计。", "reproductionZh": "除扭矩/姿态约束外还应独立建模电池与驱动器电流、母线电压和热限值；遵守原始测试保护条件，不能把论文当安全认证。", "experimentType": "both", "robots": [ "Solo-12" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "先学带特权地形信息的受约束策略，再用其经验初始化离策略视觉强化学习；让视觉策略适应有限观测，同时保留关节和姿态约束。" }, { "section": "原文实验与结果相关选段", "note": "Isaac Gym训练后部署Solo-12，Raspberry Pi 5上50Hz策略、D405深度相机；比较DAgger/BC蒸馏并统计地形完成与约束违反。 实机可爬40厘米台阶、跨35厘米缝隙并从20厘米障碍下穿过；无约束视觉RL虽可能走得更远，却更频繁违反硬件约束。" }, { "section": "局限与复现条件", "note": "作者报告实机可能触及未建模的电池电流上限，并怀疑试验中损伤电池；满足电机限制不足以保证供电安全，训练地形仍需手工设计。" } ], "originalSha256": "f903a60fc167acc21a4a99dc010bbf51634561ac6a278375001e5548cd0c17bd", "readingScopeSha256": "73f965eeeea7a81b85a65ffbccb6dc2580491e015feae7f2cb350f72468448ca", "textExtractionSha256": "80e2f6d2621757b4e78e3a8e125c8574254a676ebb8de701ec09ade0c38e2ef5" }, "experimentNote": "Isaac Gym训练后部署Solo-12，Raspberry Pi 5上50Hz策略、D405深度相机；比较DAgger/BC蒸馏并统计地形完成与约束违反。", "contribution": "先学带特权地形信息的受约束策略，再用其经验初始化离策略视觉强化学习；让视觉策略适应有限观测，同时保留关节和姿态约束。", "whyUseful": "除扭矩/姿态约束外还应独立建模电池与驱动器电流、母线电压和热限值；遵守原始测试保护条件，不能把论文当安全认证。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "根许可证 BSD-2-Clause；特定文件按头部许可，Isaac Gym 等另有条款。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2408.14035", "title": "FAST-LIVO2: Fast, Direct LiDAR-Inertial-Visual Odometry", "date": "2024-08-26", "url": "https://arxiv.org/abs/2408.14035v2", "titleZh": "FAST-LIVO2：快速直接激光雷达-惯性-视觉里程计", "abstractZh": "顺序融合激光雷达、惯性与图像观测，统一体素地图管理几何和图像块；估计曝光变化并按需光线投射弥补雷达近距盲区。\n比较公开与自采多传感器序列，并开展四组全机载无人机闭环避障飞行；机载NUC接收里程计并与规划、MPC控制集成。", "summary": "顺序融合激光雷达、惯性与图像观测，统一体素地图管理几何和图像块；估计曝光变化并按需光线投射弥补雷达近距盲区。", "experimentType": "real", "robots": [ "自主四旋翼（整机型号未核实）", "DJI M300 RTK（数据采集平台）" ], "tags": [ "导航与建图" ], "limitations": "本身是里程计，长距离仍可能漂移，原文将回环与滑窗优化列为后续方向；不能等同全局无漂移SLAM。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的直接多传感器里程计研究；不因较新年份或实机演示归为基础方法。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2408.14035v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2408.14035v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/hku-mars/FAST-LIVO2", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/hku-mars/FAST-LIVO2", "category": "导航与建图", "year": 2024, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "传统估计系统，无策略训练或基础模型预训练要求。 官方 FAST-LIVO2 数据集链接可用；另链 Global-LVBA 数据说明。 不需要神经策略检查点。 不要将关联硬件仓库、数据或设备驱动自动视为相同许可；公开代码不等于特定整机即插即用。 本次未运行训练、复现实验或实机控制。", "directions": [ "导航与建图" ], "robotFilters": [ "自主四旋翼（整机型号未核实）", "DJI M300 RTK（数据采集平台）" ], "original": { "id": "arxiv-2408.14035", "originalSourceUrl": "https://arxiv.org/pdf/2408.14035v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2408.14035v2", "finalDownloadUrl": "https://arxiv.org/pdf/2408.14035v2", "sha256": "8a9e5643ff99e4d9b735507ddd80fddb6af38401f635d72a0bf017eda57c5695", "bytes": 39653122, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:43.270255+00:00", "publicationDate": "2024/08/26", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2408.14035v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2408.14035v2", "pages": 30, "sourceIdentityStatus": "title_match", "sourceTitle": "FAST-LIVO2: Fast, Direct LiDAR-Inertial-Visual Odometry", "sourceVersion": "2408.14035v2", "sourceVersionDate": "2024/08/28", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2408.14035v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2408.14035" }, "originalAnalysis": { "id": "arxiv-2408.14035", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2408.14035v2", "sourceVersion": "2408.14035v2", "sectionsRead": [ "PDF第2页选段（提取文本L83–100）", "PDF第6页选段（提取文本L607–624）", "PDF第10页选段（提取文本L1306–1323）", "PDF第11页选段（提取文本L1450–1463）", "PDF第13页选段（提取文本L1961–1974）", "PDF第17页选段（提取文本L2438–2476）", "PDF第19页选段（提取文本L2655–2690）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "顺序融合激光雷达、惯性与图像观测，统一体素地图管理几何和图像块；估计曝光变化并按需光线投射弥补雷达近距盲区。", "experimentsZh": "比较公开与自采多传感器序列，并开展四组全机载无人机闭环避障飞行；机载NUC接收里程计并与规划、MPC控制集成。", "resultsZh": "多源约束改善退化环境鲁棒性；在低质量图像序列中视觉融合不一定增加精度，闭环飞行结果属于完整系统表现。", "limitationsZh": "本身是里程计，长距离仍可能漂移，原文将回环与滑窗优化列为后续方向；不能等同全局无漂移SLAM。", "reproductionZh": "核实LiDAR/相机/IMU时间同步、外参、曝光与噪声参数，分开评估定位算法和飞控规划器，并报告退化与失败序列。", "experimentType": "real", "robots": [ "自主四旋翼（整机型号未核实）", "DJI M300 RTK（数据采集平台）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "顺序融合激光雷达、惯性与图像观测，统一体素地图管理几何和图像块；估计曝光变化并按需光线投射弥补雷达近距盲区。" }, { "section": "原文实验与结果相关选段", "note": "比较公开与自采多传感器序列，并开展四组全机载无人机闭环避障飞行；机载NUC接收里程计并与规划、MPC控制集成。 多源约束改善退化环境鲁棒性；在低质量图像序列中视觉融合不一定增加精度，闭环飞行结果属于完整系统表现。" }, { "section": "局限与复现条件", "note": "本身是里程计，长距离仍可能漂移，原文将回环与滑窗优化列为后续方向；不能等同全局无漂移SLAM。" } ], "originalSha256": "8a9e5643ff99e4d9b735507ddd80fddb6af38401f635d72a0bf017eda57c5695", "readingScopeSha256": "e3a3ef88fee4c30af3976dc3545e0daaaba598ac21a4b2900adf585772681c8e", "textExtractionSha256": "78e31bcc76c5a7859aa7e516f7dffccc560ec6aab77e1677fa513e494a1d51e0" }, "experimentNote": "比较公开与自采多传感器序列，并开展四组全机载无人机闭环避障飞行；机载NUC接收里程计并与规划、MPC控制集成。", "contribution": "顺序融合激光雷达、惯性与图像观测，统一体素地图管理几何和图像块；估计曝光变化并按需光线投射弥补雷达近距盲区。", "whyUseful": "核实LiDAR/相机/IMU时间同步、外参、曝光与噪声参数，分开评估定位算法和飞控规划器，并报告退化与失败序列。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "代码 GPLv2；作者另提供商业替代许可联系渠道。GPL 不等于禁止商业使用，需遵守分发义务。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2406.10454", "title": "HumanPlus: Humanoid Shadowing and Imitation from Humans", "titleZh": "HumanPlus：人形机器人跟随人类动作并学习自主技能", "shortTitle": "HumanPlus", "date": "2024-06-15", "datePrecision": "day", "year": 2024, "url": "https://arxiv.org/abs/2406.10454", "paperUrl": "https://arxiv.org/abs/2406.10454", "codeUrl": "https://github.com/MarkFzp/humanplus", "summary": "HumanPlus先把AMASS人体动作重定向到H1形态，在域随机化仿真中用PPO训练带历史的全身跟踪Transformer；单RGB相机实时估计人手与身体，控制机器人采集真实双目观察和动作。HIT再以行为克隆…", "abstractZh": "HumanPlus先把AMASS人体动作重定向到H1形态，在域随机化仿真中用PPO训练带历史的全身跟踪Transformer；单RGB相机实时估计人手与身体，控制机器人采集真实双目观察和动作。HIT再以行为克隆学习50步动作段，并加未来图像特征预测，抑制只依赖本体状态而忽视视觉。\n硬件为改装Unitree H1，加入两只Inspire RH56DFX及自制腕关节，共33自由度。真人跟随与自主模仿分开评价：六名用户做遥操作对比，自主任务每项25–40示范、十次测试，对照单目、ACT和开环回放。", "category": "人形机器人 / 全身模仿学习", "tags": [ "人类动作", "行为克隆", "双目视觉", "全身操作" ], "directions": [ "人形机器人", "操作与抓取", "运动控制", "模仿学习" ], "tier": "classic", "experimentType": "both", "robots": [ "Unitree H1（加装Inspire-Robots RH56DFX及腕关节）" ], "robotNote": "", "codeStatus": "open", "status": "官方HST、HIT训练及硬件代码公开。", "trainingNote": "先训练Humanoid Shadowing Transformer，再用真实遥操作数据训练Humanoid Imitation Transformer；需专门硬件与动作数据。", "contribution": "HumanPlus先把AMASS人体动作重定向到H1形态，在域随机化仿真中用PPO训练带历史的全身跟踪Transformer；单RGB相机实时估计人手与身体，控制机器人采集真实双目观察和动作。HIT再以行为克隆学习50步动作段，并加未来图像特征预测，抑制只依赖本体状态而忽视视觉。", "whyUseful": "复现需改装硬件、固定姿态映射、AMASS处理、域随机化和双目同步。低层50Hz、PD1000Hz，人体/手估计约25/10fps；自主策略数据来自真机跟随，不能归纳为全流程仅靠仿真人体视频迁移。", "limitations": "每项仅十次，不能说明开放场景可靠性。四自由度手臂加腕仍不能覆盖任意末端姿态，固定相机存在视野盲区，遮挡影响人体估计；坐姿时绕过学习控制直接发关节目标，长程导航仍未解决。", "evidence": [ { "url": "https://arxiv.org/abs/2406.10454", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/html/2406.10454v1", "note": "§3明确改装H1、RH56DFX手与定制腕；任务与实验分开报告跟随和自主模仿。" }, { "url": "https://humanoid-ai.github.io/", "note": "作者项目页、数据和硬件入口。" }, { "url": "https://github.com/MarkFzp/humanplus", "note": "官方仓库含HST与HIT训练命令和硬件代码。" }, { "url": "https://arxiv.org/html/2406.10454v1", "note": "3–5：33DoF改装H1；仿真低层，真实跟随数据，高层模仿。" }, { "url": "https://arxiv.org/html/2406.10454v1", "note": "9; Table 5：十次/任务；各阶段分母为到达阶段次数。" }, { "url": "https://arxiv.org/html/2406.10454v1", "note": "5; 10：坐姿直接PD；遮挡、视野和自由度限制。" }, { "url": "https://arxiv.org/html/2406.10454v1", "note": "PDF p9, Table 5：已渲染完整表，六任务整段成功率100/90/80/90/90/60；阶段列分母随前序成功减少。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2024：人类动作先验成为人形遥操作采集与自主模仿学习的数据入口。", "projectUrl": null, "robotFilters": [ "Unitree H1" ], "original": { "id": "arxiv-2406.10454", "originalSourceUrl": "https://arxiv.org/pdf/2406.10454v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2406.10454v1", "finalDownloadUrl": "https://arxiv.org/pdf/2406.10454", "sha256": "94c5fe16631e2d1c5b5bdd8326adda17c719ad63b2630979ca6aa9e0a7af4dd2", "bytes": 3711077, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:37:00.889735+00:00", "publicationDate": "2024/06/15", "publicationDateSourceUrl": "https://arxiv.org/abs/2406.10454", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2406.10454", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2406.10454", "pages": 17, "sourceIdentityStatus": "title_match", "sourceTitle": "HumanPlus: Humanoid Shadowing and Imitation from Humans", "sourceVersion": "2406.10454v1", "sourceVersionDate": "2024/06/15", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.953855+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2406.10454" }, "originalAnalysis": { "id": "arxiv-2406.10454", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:45:34.458249+00:00", "sourceUrl": "https://arxiv.org/html/2406.10454v1", "sectionsRead": [ "3 Hardware", "4 Human Data", "5 Shadowing", "6 Imitation", "8–9 Experiments", "Table 5", "10 Limitations" ], "methodsZh": "HumanPlus先把AMASS人体动作重定向到H1形态，在域随机化仿真中用PPO训练带历史的全身跟踪Transformer；单RGB相机实时估计人手与身体，控制机器人采集真实双目观察和动作。HIT再以行为克隆学习50步动作段，并加未来图像特征预测，抑制只依赖本体状态而忽视视觉。", "experimentsZh": "硬件为改装Unitree H1，加入两只Inspire RH56DFX及自制腕关节，共33自由度。真人跟随与自主模仿分开评价：六名用户做遥操作对比，自主任务每项25–40示范、十次测试，对照单目、ACT和开环回放。", "resultsZh": "作者报告自主折衣100%、重排/握手/仓储90%、打字80%，穿鞋站起走路60%；子任务百分比以到达该阶段的尝试数为分母。扰动恢复相对原厂控制器更快，但弹琴、乒乓、拳击等是人类跟随展示，不是同等自主技能结果。", "limitationsZh": "每项仅十次，不能说明开放场景可靠性。四自由度手臂加腕仍不能覆盖任意末端姿态，固定相机存在视野盲区，遮挡影响人体估计；坐姿时绕过学习控制直接发关节目标，长程导航仍未解决。", "reproductionZh": "复现需改装硬件、固定姿态映射、AMASS处理、域随机化和双目同步。低层50Hz、PD1000Hz，人体/手估计约25/10fps；自主策略数据来自真机跟随，不能归纳为全流程仅靠仿真人体视频迁移。", "experimentType": "both", "robots": [ "Unitree H1（加装Inspire-Robots RH56DFX及腕关节）" ], "evidenceNotes": [ { "section": "3–5", "note": "33DoF改装H1；仿真低层，真实跟随数据，高层模仿。" }, { "section": "9; Table 5", "note": "十次/任务；各阶段分母为到达阶段次数。" }, { "section": "5; 10", "note": "坐姿直接PD；遮挡、视野和自由度限制。" }, { "section": "PDF p9, Table 5", "note": "已渲染完整表，六任务整段成功率100/90/80/90/90/60；阶段列分母随前序成功减少。" } ], "corrections": [ "明确跟随展示与自主模仿任务；不是所有视频技能均自主。" ], "sourceVersion": "2406.10454v1", "originalSha256": "94c5fe16631e2d1c5b5bdd8326adda17c719ad63b2630979ca6aa9e0a7af4dd2" }, "experimentNote": "硬件为改装Unitree H1，加入两只Inspire RH56DFX及自制腕关节，共33自由度。真人跟随与自主模仿分开评价：六名用户做遥操作对比，自主任务每项25–40示范、十次测试，对照单目、ACT和开环回放。", "analysisVerifiedAt": "2026-10-04T13:45:34.458249+00:00" }, { "id": "arxiv-2406.08858", "title": "OmniH2O: Universal and Dexterous Human-to-Humanoid Whole-Body Teleoperation and Learning", "titleZh": "OmniH2O：通用灵巧的人到人形全身遥操作与学习", "shortTitle": "OmniH2O", "date": "2024-06-13", "datePrecision": "day", "year": 2024, "url": "https://arxiv.org/abs/2406.08858", "paperUrl": "https://arxiv.org/abs/2406.08858", "codeUrl": "https://github.com/LeCAR-Lab/human2humanoid", "summary": "先重定向并筛选人体动作，以全状态特权信息训练全身模仿教师，再用DAgger蒸馏成只依赖头/双手三点目标和25步本体历史的学生。学生无需显式全局线速度估计；灵巧手由VR姿态经逆运动学控制，另以视觉模仿策略输出运动…", "abstractZh": "先重定向并筛选人体动作，以全状态特权信息训练全身模仿教师，再用DAgger蒸馏成只依赖头/双手三点目标和25步本体历史的学生。学生无需显式全局线速度估计；灵巧手由VR姿态经逆运动学控制，另以视觉模仿策略输出运动目标。\n仿真在约1.4万条重定向增强动作上比较教师、H2O及历史/关键点消融；真实Unitree H1加Inspire手评测20条站立动作，并演示VR、相机、语言接口。收集约40分钟六任务示范，四任务各十次评测自主模仿。", "category": "人形机器人 / 全身操作", "tags": [ "遥操作", "教师学生", "稀疏输入", "自主技能" ], "directions": [ "人形机器人", "运动控制", "操作与抓取", "灵巧手" ], "tier": "classic", "experimentType": "both", "robots": [ "Unitree H1", "Inspire手" ], "robotNote": "论文只给出Inspire手品牌，未给精确手部子型号；不借用HumanPlus的RH56DFX型号。", "codeStatus": "open", "status": "官方教师训练、学生蒸馏与动作资源公开；许可限制非商业用途。", "trainingNote": "通过RL训练特权教师，再进行基于历史的DAgger式学生蒸馏；与H2O共用代码库并有独立配置。", "contribution": "先重定向并筛选人体动作，以全状态特权信息训练全身模仿教师，再用DAgger蒸馏成只依赖头/双手三点目标和25步本体历史的学生。学生无需显式全局线速度估计；灵巧手由VR姿态经逆运动学控制，另以视觉模仿策略输出运动目标。", "whyUseful": "应记录人体到H1重定向、稳定站立增强、奖励课程和域随机化，并分别校验50Hz策略与200HzPD接口。比较时对齐全局/相对误差和成功阈值；未实机复现。", "limitations": "大规模动作覆盖来自仿真，真实量化主要为站立序列；三点输入可能对应多个全身姿态。灵巧手直接映射和全身策略是不同模块，GPT-4o演示选运动原语不等于通用视觉闭环操作。", "evidence": [ { "url": "https://arxiv.org/abs/2406.08858", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/html/2406.08858v1", "note": "图1及附录A明确Unitree H1、Inspire手和真实计算配置；附录说明指腕直接映射。" }, { "url": "https://omni.human2humanoid.com/", "note": "作者页面分列遥操作、GPT-4o与Diffusion Policy自主展示。" }, { "url": "https://github.com/LeCAR-Lab/human2humanoid", "note": "官方教师训练、学生蒸馏配置及CC BY-NC许可。" }, { "url": "https://arxiv.org/pdf/2406.08858", "note": "§3; Figure 3：特权教师、DAgger、25步历史和三点目标。" }, { "url": "https://arxiv.org/pdf/2406.08858", "note": "Tables 1–3; Appendix A：实机H1及Inspire手、20站立序列、四任务自主平均。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2024：从全身遥操作扩展到统一输入接口和示范驱动的自主技能。", "projectUrl": null, "robotFilters": [ "Unitree H1", "Inspire手" ], "original": { "id": "arxiv-2406.08858", "originalSourceUrl": "https://arxiv.org/pdf/2406.08858v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2406.08858v1", "finalDownloadUrl": "https://arxiv.org/pdf/2406.08858", "sha256": "bedce4c95c9776d64cd29587029c1f4d8171ad20f45017493627b27eb6b3138e", "bytes": 8193229, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:37:24.379623+00:00", "publicationDate": "2024/06/13", "publicationDateSourceUrl": "https://arxiv.org/abs/2406.08858", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2406.08858", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2406.08858", "pages": 25, "sourceIdentityStatus": "title_match", "sourceTitle": "OmniH2O: Universal and Dexterous Human-to-Humanoid Whole-Body Teleoperation and Learning", "sourceVersion": "2406.08858v1", "sourceVersionDate": "2024/06/13", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.972660+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2406.08858" }, "originalAnalysis": { "id": "arxiv-2406.08858", "sourceUrl": "https://arxiv.org/pdf/2406.08858", "sectionsRead": [ "§3", "§4.1–4.3", "Tables 1–3", "Appendix A" ], "methodsZh": "先重定向并筛选人体动作，以全状态特权信息训练全身模仿教师，再用DAgger蒸馏成只依赖头/双手三点目标和25步本体历史的学生。学生无需显式全局线速度估计；灵巧手由VR姿态经逆运动学控制，另以视觉模仿策略输出运动目标。", "experimentsZh": "仿真在约1.4万条重定向增强动作上比较教师、H2O及历史/关键点消融；真实Unitree H1加Inspire手评测20条站立动作，并演示VR、相机、语言接口。收集约40分钟六任务示范，四任务各十次评测自主模仿。", "resultsZh": "作者报告仿真学生追踪成功94.10%，教师94.77%、H2O87.52%；真实全局关节位置误差47.94毫米，对照87.33毫米。四个自主任务扩散策略平均约8/10，普通BC约1/10；这些是所选任务平均，不代表所有遥操作技能自主化。", "limitationsZh": "大规模动作覆盖来自仿真，真实量化主要为站立序列；三点输入可能对应多个全身姿态。灵巧手直接映射和全身策略是不同模块，GPT-4o演示选运动原语不等于通用视觉闭环操作。", "reproductionZh": "应记录人体到H1重定向、稳定站立增强、奖励课程和域随机化，并分别校验50Hz策略与200HzPD接口。比较时对齐全局/相对误差和成功阈值；未实机复现。", "experimentType": "both", "robots": [ "Unitree H1", "Inspire手" ], "evidenceNotes": [ { "section/page": "§3; Figure 3", "note": "特权教师、DAgger、25步历史和三点目标。" }, { "section/page": "Tables 1–3; Appendix A", "note": "实机H1及Inspire手、20站立序列、四任务自主平均。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "仿真在约1.4万条重定向增强动作上比较教师、H2O及历史/关键点消融；真实Unitree H1加Inspire手评测20条站立动作，并演示VR、相机、语言接口。收集约40分钟六任务示范，四任务各十次评测自主模仿。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2406.09246", "title": "OpenVLA: An Open-Source Vision-Language-Action Model", "titleZh": "OpenVLA：开源视觉－语言－动作模型", "shortTitle": "OpenVLA", "date": "2024-06-13", "year": 2024, "url": "https://arxiv.org/abs/2406.09246", "codeUrl": "https://github.com/openvla/openvla", "category": "视觉语言动作", "tags": [ "VLA", "LoRA", "量化", "开源" ], "tier": "foundation", "summary": "OpenVLA将DINOv2与SigLIP特征拼接到Llama2/Prismatic7B骨干，把每维连续动作离散为256分箱并自回归预测。用筛选后的Open X-Embodiment约97万轨迹微调全部模块，支…", "abstractZh": "OpenVLA将DINOv2与SigLIP特征拼接到Llama2/Prismatic7B骨干，把每维连续动作离散为256分箱并自回归预测。用筛选后的Open X-Embodiment约97万轨迹微调全部模块，支持LoRA适配与低比特推理。\n直接评估WidowX17任务×10次、Google移动操作平台12任务×5次，按相同初态A/B比较RT-1-X、RT-2-X和Octo。另在两种Panda设置用每任务10–150示范适配七任务，共129测试回合，并测33回合LoRA与80回合量化。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "real", "robots": [ "WidowX", "Google移动操作机器人", "Franka Emika Panda" ], "codeStatus": "open", "status": "官方训练、微调、评测代码及模型权重公开。", "trainingNote": "代码 MIT；权重及基础模型仍有独立许可。预训练约 97 万真实轨迹；LIBERO 是后加的仿真微调评测。", "whyUseful": "完整预训练为64张A100约14天；实际复现宜先验证公开检查点、动作归一化和低成本微调。需匹配控制频率，八比特实验降速会改变闭环动力学；本轮核读论文流程但未训练或实机运行。", "contribution": "OpenVLA将DINOv2与SigLIP特征拼接到Llama2/Prismatic7B骨干，把每维连续动作离散为256分箱并自回归预测。用筛选后的Open X-Embodiment约97万轨迹微调全部模块，支持LoRA适配与低比特推理。", "limitations": "原版只用单图、无历史及本体输入，推理频率限制精细高频控制，窄域灵巧任务Diffusion Policy更平滑。LoRA/量化表使用较小训练混合及仅SigLIP变体，不应当全部主模型的无条件保证。", "caveats": "日期为 v1 首发；实验范围采用含 LIBERO 附录的 2024-09-05 v3。", "evidence": [ { "url": "https://arxiv.org/abs/2406.09246", "note": "首发日期、参数规模与摘要。" }, { "url": "https://arxiv.org/html/2406.09246v3", "note": "第 5 节及附录 E 区分实机和 LIBERO 仿真。" }, { "url": "https://github.com/openvla/openvla", "note": "MIT、实际代码、检查点和 2024-09 仿真实验更新。" }, { "url": "https://arxiv.org/html/2406.09246v1", "note": "§5.1–5.2：直接评估与Franka适配的任务/回合数。" }, { "url": "https://arxiv.org/html/2406.09246v1", "note": "§5.3 footnote 4：LoRA/量化实验是较小数据、SigLIP-only变体。" }, { "url": "https://arxiv.org/html/2406.09246v1", "note": "§6：单图、频率与通常低于90%可靠性的限制。" } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "视觉语言动作" ], "robotFilters": [ "WidowX", "Google Robot", "Franka Panda" ], "verificationNote": "已核 v3 论文与官方仓库；与后续 OpenVLA-OFT 区分。", "original": { "id": "arxiv-2406.09246", "originalSourceUrl": "https://arxiv.org/pdf/2406.09246v3", "versionedOriginalUrl": "https://arxiv.org/pdf/2406.09246v3", "finalDownloadUrl": "https://arxiv.org/pdf/2406.09246", "sha256": "353c37df34458f12f969b14dfd8b77175b727b9cddea7bb891759beddeefe1be", "bytes": 13240309, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:37:01.082684+00:00", "publicationDate": "2024/06/13", "publicationDateSourceUrl": "https://arxiv.org/abs/2406.09246", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2406.09246", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2406.09246", "pages": 37, "sourceIdentityStatus": "title_match", "sourceTitle": "OpenVLA: An Open-Source Vision-Language-Action Model", "sourceVersion": "2406.09246v3", "sourceVersionDate": "2024/09/05", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.964446+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2406.09246" }, "originalAnalysis": { "id": "arxiv-2406.09246", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090132+00:00", "sourceUrl": "https://arxiv.org/html/2406.09246v1", "sectionsRead": [ "§3–4", "§5.1–5.4", "Tables 1–2", "§6" ], "methodsZh": "OpenVLA将DINOv2与SigLIP特征拼接到Llama2/Prismatic7B骨干，把每维连续动作离散为256分箱并自回归预测。用筛选后的Open X-Embodiment约97万轨迹微调全部模块，支持LoRA适配与低比特推理。", "experimentsZh": "直接评估WidowX17任务×10次、Google移动操作平台12任务×5次，按相同初态A/B比较RT-1-X、RT-2-X和Octo。另在两种Panda设置用每任务10–150示范适配七任务，共129测试回合，并测33回合LoRA与80回合量化。", "resultsZh": "作者报告WidowX总体超过RT-2-X，Google平台相近，但语义泛化RT-2-X仍更强。选定微调任务LoRA为68.2±7.5%，全微调69.7±7.2%；四比特示例约7GB显存且表现接近半精度。", "limitationsZh": "原版只用单图、无历史及本体输入，推理频率限制精细高频控制，窄域灵巧任务Diffusion Policy更平滑。LoRA/量化表使用较小训练混合及仅SigLIP变体，不应当全部主模型的无条件保证。", "reproductionZh": "完整预训练为64张A100约14天；实际复现宜先验证公开检查点、动作归一化和低成本微调。需匹配控制频率，八比特实验降速会改变闭环动力学；本轮核读论文流程但未训练或实机运行。", "experimentType": "real", "robots": [ "WidowX", "Google移动操作机器人", "Franka Emika Panda" ], "evidenceNotes": [ { "section/page": "§5.1–5.2", "note": "直接评估与Franka适配的任务/回合数。" }, { "section/page": "§5.3 footnote 4", "note": "LoRA/量化实验是较小数据、SigLIP-only变体。" }, { "section/page": "§6", "note": "单图、频率与通常低于90%可靠性的限制。" } ] }, "experimentNote": "直接评估WidowX17任务×10次、Google移动操作平台12任务×5次，按相同初态A/B比较RT-1-X、RT-2-X和Octo。另在两种Panda设置用每任务10–150示范适配七任务，共129测试回合，并测33回合LoRA与80回合量化。", "analysisVerifiedAt": "2026-10-04T13:50:35.090132+00:00" }, { "id": "arxiv-2406.02523", "title": "RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots", "titleZh": "RoboCasa：面向通用机器人的大规模日常任务仿真", "shortTitle": "RoboCasa (2024)", "date": "2024-06-04", "datePrecision": "day", "year": 2024, "url": "https://arxiv.org/abs/2406.02523", "paperUrl": "https://arxiv.org/abs/2406.02523", "codeUrl": "https://github.com/robocasa/robocasa", "summary": "RoboCasa把厨房布局、装置、对象与任务组织为MuJoCo仿真框架，以生成纹理和三维资产扩大外观多样性；人类少量种子示范经MimicGen适配为大量新场景轨迹，支持原子技能、组合任务及多任务行为克隆。", "abstractZh": "RoboCasa把厨房布局、装置、对象与任务组织为MuJoCo仿真框架，以生成纹理和三维资产扩大外观多样性；人类少量种子示范经MimicGen适配为大量新场景轨迹，支持原子技能、组合任务及多任务行为克隆。\n初版包含120场景、2500多物体与100任务；比较50条/任务人类数据和100、1000、3000条生成数据。仿真主平台Panda加Omron底盘，五评测厨房含未见风格和对象；真实DROID配置Panda做三个厨房抓放任务。", "category": "数据集与基准 / 家庭操作", "tags": [ "家庭操作", "合成数据", "模仿学习", "多任务" ], "directions": [ "数据集与基准", "操作与抓取", "模仿学习" ], "tier": "classic", "experimentType": "both", "robots": [ "Franka Emika Panda", "Omron移动底盘（仿真）" ], "robotNote": "原文仿真主平台为Panda + Omron移动底座；真实任务为Panda机械臂，本条不推填实机底座型号。", "codeStatus": "open", "status": "官方环境、数据与策略训练支持公开；需选择与2024论文对应的历史版本。", "trainingNote": "原文使用robomimic BC-Transformer与MimicGen生成示范；当前仓库已扩展为RoboCasa365，复现应对照旧文档和数据。", "contribution": "RoboCasa把厨房布局、装置、对象与任务组织为MuJoCo仿真框架，以生成纹理和三维资产扩大外观多样性；人类少量种子示范经MimicGen适配为大量新场景轨迹，支持原子技能、组合任务及多任务行为克隆。", "whyUseful": "固定生成轨迹总量、原子任务覆盖及训练/评估纹理差异；先复现BC-Transformer规模曲线，再做真实加仿真的同数据量对照。报告组合任务分项，筛除机械执行不良的成功轨迹。", "limitations": "合成“成功”轨迹可含碰撞和抖动，任务代码仍需人工。真实绝对成功率低，不能把数据增益说成已解决家务；仿真20Hz和实机15Hz、控制器及标定都有差异，场景仍集中厨房。", "evidence": [ { "url": "https://arxiv.org/abs/2406.02523", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/html/2406.02523v1", "note": "§V说明Panda仿真设置与三个真实Panda任务；结论明确原版100任务与局限。" }, { "url": "https://github.com/robocasa/robocasa", "note": "官方仓库区分2024原版与2026 RoboCasa365，提供代码与版本记录。" }, { "url": "https://robocasa.ai/docs/build/html/v0.2/use_cases/policy_learning.html", "note": "历史官方文档说明robomimic分支与BC-Transformer训练。" }, { "url": "https://arxiv.org/pdf/2406.02523", "note": "III–IV; V-A：场景资产规模、MimicGen及Panda/Omron。" }, { "url": "https://arxiv.org/pdf/2406.02523", "note": "V-C Figs 9–10; VI：真实Panda、成功率提升及合成轨迹质量限制。" }, { "url": "https://arxiv.org/pdf/2406.02523v1", "note": "III–IV; V-A：场景资产规模、MimicGen及Panda/Omron。" }, { "url": "https://arxiv.org/pdf/2406.02523v1", "note": "V-C Figs 9–10; VI：真实Panda、成功率提升及合成轨迹质量限制。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2024：从桌面单任务基准走向多样厨房与可扩展日常操作数据。", "projectUrl": null, "robotFilters": [ "Franka Panda" ], "original": { "id": "arxiv-2406.02523", "originalSourceUrl": "https://arxiv.org/pdf/2406.02523v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2406.02523v1", "finalDownloadUrl": "https://arxiv.org/pdf/2406.02523", "sha256": "7aca93ca3f157e32a46e02d675b3b2c24e3073a7807c038f7f566b73db87bde3", "bytes": 8307032, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:37:29.645195+00:00", "publicationDate": "2024/06/04", "publicationDateSourceUrl": "https://arxiv.org/abs/2406.02523", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2406.02523", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2406.02523", "pages": 16, "sourceIdentityStatus": "title_match", "sourceTitle": "RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots", "sourceVersion": "2406.02523v1", "sourceVersionDate": "2024/06/04", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.979456+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2406.02523" }, "originalAnalysis": { "id": "arxiv-2406.02523", "sourceUrl": "https://arxiv.org/pdf/2406.02523v1", "sectionsRead": [ "III", "IV", "V-A–C", "VI", "Figs 7–10" ], "methodsZh": "RoboCasa把厨房布局、装置、对象与任务组织为MuJoCo仿真框架，以生成纹理和三维资产扩大外观多样性；人类少量种子示范经MimicGen适配为大量新场景轨迹，支持原子技能、组合任务及多任务行为克隆。", "experimentsZh": "初版包含120场景、2500多物体与100任务；比较50条/任务人类数据和100、1000、3000条生成数据。仿真主平台Panda加Omron底盘，五评测厨房含未见风格和对象；真实DROID配置Panda做三个厨房抓放任务。", "resultsZh": "原子技能平均成功由人类数据28.8%增到完整生成数据47.6%；组合任务仍困难，预训练微调在五任务中四项非零。实机联合仿真数据后，已见对象平均13.6%到24.4%，未见对象2.6%到9.3%，是有限但明确的跨域收益。", "limitationsZh": "合成“成功”轨迹可含碰撞和抖动，任务代码仍需人工。真实绝对成功率低，不能把数据增益说成已解决家务；仿真20Hz和实机15Hz、控制器及标定都有差异，场景仍集中厨房。", "reproductionZh": "固定生成轨迹总量、原子任务覆盖及训练/评估纹理差异；先复现BC-Transformer规模曲线，再做真实加仿真的同数据量对照。报告组合任务分项，筛除机械执行不良的成功轨迹。", "experimentType": "both", "robots": [ "Franka Emika Panda", "Omron移动底盘（仿真）" ], "corrections": [ "原论文不只有仿真，还包含低成功率的真实厨房联合训练实验。" ], "evidenceNotes": [ { "note": "场景资产规模、MimicGen及Panda/Omron。", "section": "III–IV; V-A" }, { "note": "真实Panda、成功率提升及合成轨迹质量限制。", "section": "V-C Figs 9–10; VI" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2406.02523v1", "originalSha256": "7aca93ca3f157e32a46e02d675b3b2c24e3073a7807c038f7f566b73db87bde3" }, "experimentNote": "初版包含120场景、2500多物体与100任务；比较50条/任务人类数据和100、1000、3000条生成数据。仿真主平台Panda加Omron底盘，五评测厨房含未见风格和对象；真实DROID配置Panda做三个厨房抓放任务。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2405.12213", "title": "Octo: An Open-Source Generalist Robot Policy", "titleZh": "Octo：开源通用机器人策略", "shortTitle": "Octo", "date": "2024-05-20", "year": 2024, "url": "https://arxiv.org/abs/2405.12213", "codeUrl": "https://github.com/octo-models/octo", "category": "机器人基础模型", "tags": [ "扩散策略", "跨本体", "微调", "开源" ], "tier": "foundation", "summary": "将语言、目标图像和多相机历史编码成分块注意力Transformer输入，以只读取上下文的读出令牌连接扩散动作块解码器。增减传感器或动作头时保留主体权重；从Open X-Embodiment筛选25套数据、约80…", "abstractZh": "将语言、目标图像和多相机历史编码成分块注意力Transformer输入，以只读取上下文的读出令牌连接扩散动作块解码器。增减传感器或动作头时保留主体权重；从Open X-Embodiment筛选25套数据、约80万条轨迹，统一末端增量及夹爪语义。\n四家机构九套真实系统分别检验训练域内零样本控制与新域微调。六个微调域约用100条示范，覆盖力矩输入、关节动作、ViperX和ALOHA；对照从零训练与VC-1。零样本每机器人两任务、每任务10次；微调表I标为每域20次，但双臂附录写10次，原文存在协议差异。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "real", "robots": [ "Trossen WidowX 250", "UR5", "RT-1 Robot（专有平台）", "Franka", "Trossen ViperX", "ALOHA（双ViperX）" ], "codeStatus": "open", "status": "MIT 训练、微调与推理代码和预训练权重公开。", "trainingNote": "约 80 万条 OXE 轨迹；需区分原论文模型与仓库后续 1.5 检查点。模型头适配不等于免示范迁移。", "whyUseful": "论文给出JAX预训练/微调及27M、93M权重；Base用128芯TPU v4训练14小时，24GB A5000微调约5小时。复现须固定数据混合、动作归一化、相机及逐任务成功判据，不能只加载检查点。", "contribution": "将语言、目标图像和多相机历史编码成分块注意力Transformer输入，以只读取上下文的读出令牌连接扩散动作块解码器。增减传感器或动作头时保留主体权重；从Open X-Embodiment筛选25套数据、约80万条轨迹，统一末端增量及夹爪语义。", "limitations": "样本量有限，跨模型预训练数据量不一致；依赖末端动作筛选的数据配方，未提供通用安全保证。九套系统也不等于九种互不重复的机器人型号。", "caveats": "论文称 9 套实机配置，不应写成 9 个不同商业机器人型号。", "evidence": [ { "url": "https://arxiv.org/abs/2405.12213", "note": "首发日期与总体研究范围。" }, { "url": "https://arxiv.org/html/2405.12213v1", "note": "附录 F 列出 9 套实机配置、WidowX/UR5/ViperX/ALOHA 等。" }, { "url": "https://github.com/octo-models/octo", "note": "MIT 许可证、训练脚本和模型权重；版本 1.5 更新。" }, { "url": "https://arxiv.org/pdf/2405.12213v2", "note": "III-B–D：25 datasets/800k episodes; 300k steps, batch 2048; 50k-step full-model finetuning." }, { "url": "https://arxiv.org/pdf/2405.12213v2", "note": "表I、附录F-C：Table I says 20 trials/domain; bimanual appendix says 10. Preserve discrepancy." }, { "url": "https://arxiv.org/pdf/2405.12213v2", "note": "表VII：In-distribution 85%, novel objects 80%, new environment 40%, new skill 5%." }, { "url": "https://arxiv.org/pdf/2405.12213v2", "note": "附录F：Explicit hardware: WidowX 250, UR5, proprietary RT-1 robot, ViperX, ALOHA; CMU setup uses Franka controller." } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "其他机器人研究" ], "robotFilters": [ "WidowX 250", "Universal Robots UR5", "Franka（型号未注明）", "Trossen ViperX", "ALOHA" ], "verificationNote": "已核原论文附录与官方仓库；实验标签不把后续他人 LIBERO 评测并入原论文。", "original": { "id": "arxiv-2405.12213", "originalSourceUrl": "https://arxiv.org/pdf/2405.12213v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2405.12213v2", "finalDownloadUrl": "https://arxiv.org/pdf/2405.12213", "sha256": "73bff297cfafe523319162124e6b7f96919c0930e0a380f307255c4c7464ac93", "bytes": 3675649, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:37:29.587866+00:00", "publicationDate": "2024/05/20", "publicationDateSourceUrl": "https://arxiv.org/abs/2405.12213", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2405.12213", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2405.12213", "pages": 17, "sourceIdentityStatus": "title_match", "sourceTitle": "Octo: An Open-Source Generalist Robot Policy", "sourceVersion": "2405.12213v2", "sourceVersionDate": "2024/05/26", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.983025+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2405.12213" }, "originalAnalysis": { "id": "arxiv-2405.12213", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739375+00:00", "sourceUrl": "https://arxiv.org/pdf/2405.12213v2", "sectionsRead": [ "III-A–E 架构、数据与训练", "IV-A–C 实验与表I/II", "附录D–F 超参、失败设计、各平台协议与表VII" ], "methodsZh": "将语言、目标图像和多相机历史编码成分块注意力Transformer输入，以只读取上下文的读出令牌连接扩散动作块解码器。增减传感器或动作头时保留主体权重；从Open X-Embodiment筛选25套数据、约80万条轨迹，统一末端增量及夹爪语义。", "experimentsZh": "四家机构九套真实系统分别检验训练域内零样本控制与新域微调。六个微调域约用100条示范，覆盖力矩输入、关节动作、ViperX和ALOHA；对照从零训练与VC-1。零样本每机器人两任务、每任务10次；微调表I标为每域20次，但双臂附录写10次，原文存在协议差异。", "resultsZh": "作者表I报告微调平均成功率72%，从零训练20%、VC-1为15%；WidowX消融支持扩散头和宽数据混合。表VII中未见物体80%，未见环境40%，未见技能仅5%，不能把训练域内零样本结果当作任意新技能泛化。", "limitationsZh": "样本量有限，跨模型预训练数据量不一致；依赖末端动作筛选的数据配方，未提供通用安全保证。九套系统也不等于九种互不重复的机器人型号。", "reproductionZh": "论文给出JAX预训练/微调及27M、93M权重；Base用128芯TPU v4训练14小时，24GB A5000微调约5小时。复现须固定数据混合、动作归一化、相机及逐任务成功判据，不能只加载检查点。", "evidenceNotes": [ { "section/page": "III-B–D", "note": "25 datasets/800k episodes; 300k steps, batch 2048; 50k-step full-model finetuning." }, { "section/page": "表I、附录F-C", "note": "Table I says 20 trials/domain; bimanual appendix says 10. Preserve discrepancy." }, { "section/page": "表VII", "note": "In-distribution 85%, novel objects 80%, new environment 40%, new skill 5%." }, { "section/page": "附录F", "note": "Explicit hardware: WidowX 250, UR5, proprietary RT-1 robot, ViperX, ALOHA; CMU setup uses Franka controller." } ], "experimentType": "real", "robots": [ "Trossen WidowX 250", "UR5", "RT-1 Robot（专有平台）", "Franka", "Trossen ViperX", "ALOHA（双ViperX）" ], "corrections": [ "将九套评测系统与九种机器人型号区分。", "保留双臂试验次数在表I与附录之间的不一致。" ], "sourceVersion": "2405.12213v2", "originalSha256": "73bff297cfafe523319162124e6b7f96919c0930e0a380f307255c4c7464ac93" }, "experimentNote": "四家机构九套真实系统分别检验训练域内零样本控制与新域微调。六个微调域约用100条示范，覆盖力矩输入、关节动作、ViperX和ALOHA；对照从零训练与VC-1。零样本每机器人两任务、每任务10次；微调表I标为每域20次，但双臂附录写10次，原文存在协议差异。", "analysisVerifiedAt": "2026-10-04T13:47:58.739375+00:00" }, { "id": "arxiv-2404.12377", "title": "RoboDreamer: Learning Compositional World Models for Robot Imagination", "titleZh": "RoboDreamer：为机器人想象学习可组合世界模型", "shortTitle": "RoboDreamer", "date": "2024-04-18", "year": 2024, "url": "https://arxiv.org/abs/2404.12377", "paperUrl": "https://arxiv.org/abs/2404.12377", "projectUrl": "https://robovideo.github.io/", "codeUrl": "https://github.com/rainbow979/robodreamer", "summary": "将动作语言拆为动词和空间关系短语，组合各条件的视频扩散得分来生成未来图像计划；可再加入目标图或草图。执行时用相邻预测帧与当前状态训练逆动力学模型把视频转成动作，因而视频想象与控制是两个模块。", "abstractZh": "将动作语言拆为动词和空间关系短语，组合各条件的视频扩散得分来生成未来图像计划；可再加入目标图或草图。执行时用相邻预测帧与当前状态训练逆动力学模型把视频转成动作，因而视频想象与控制是两个模块。\n以RT-1约7万真实记录、约500任务训练视频模型，抽取未见语言组合；约128样本由至少三人评估。机器人规划在RLBench仿真Panda上使用前视RGB与宏步动作，表3列六项任务，对照Image-BC、Hiveformer、UniPi。", "category": "世界模型 / 组合式视频规划", "tags": [ "RoboDreamer", "世界模型", "强化学习", "操作与抓取" ], "directions": [ "世界模型", "操作与抓取", "强化学习" ], "tier": "recent", "experimentType": "sim", "experimentNote": "以RT-1约7万真实记录、约500任务训练视频模型，抽取未见语言组合；约128样本由至少三人评估。机器人规划在RLBench仿真Panda上使用前视RGB与宏步动作，表3列六项任务，对照Image-BC、Hiveformer、UniPi。", "robots": [ "Franka Emika Panda" ], "robotFilters": [], "robotNote": "Franka Panda 仅为 RLBench 仿真执行平台；RT-X 真实视频不计作实机部署。", "codeStatus": "unknown", "status": "代码公开，许可未核实", "trainingStatus": "代码公开，许可未核实", "trainingNote": "官方仓库含 train_rtx.py、数据管道与配置；统一许可及完整权重发布范围未核实。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "未核实", "contribution": "将动作语言拆为动词和空间关系短语，组合各条件的视频扩散得分来生成未来图像计划；可再加入目标图或草图。执行时用相邻预测帧与当前状态训练逆动力学模型把视频转成动作，因而视频想象与控制是两个模块。", "whyUseful": "复现需保留语言组合拆分、数据划分、生成分辨率和逆动力学训练，区分纯语言与额外目标图条件。附录给出约100张V100、三阶段扩散及逆模型配置；未据RLBench默认硬件推断真实Panda实验。", "limitations": "视频人评是想象计划合理性，不是实机成功率；真实RT-1数据也不构成新实体部署。论文明确单视角、真实新图像泛化不足和移动相机困难，且宏步设置弱化了低层连续控制。", "caveats": "执行验证为仿真；单相机、移动视角和未见真实图像泛化受限。", "evidence": [ { "url": "https://arxiv.org/abs/2404.12377", "note": "首发日期和标题。" }, { "url": "https://arxiv.org/html/2404.12377v1", "note": "附录 A.2 明确 RLBench Franka Panda；结论列单视角和现实域外局限。" }, { "url": "https://github.com/rainbow979/robodreamer", "note": "由官方项目页链接；含训练源码，统一许可未核实。" }, { "url": "https://arxiv.org/html/2404.12377v1", "note": "Table 3：六项RLBench均值49.3%；非实机。" }, { "url": "https://arxiv.org/html/2404.12377v1", "note": "§4.1 / Appendix A.2：真实数据用于生成评估，至少3个人评、约128样本。" }, { "url": "https://arxiv.org/html/2404.12377v1", "note": "§6：单视角、真实图像和移动相机限制。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2024：通过语言结构组合世界模型，推进未见目标的视频规划。", "freshness": "近年进展", "original": { "id": "arxiv-2404.12377", "originalSourceUrl": "https://arxiv.org/pdf/2404.12377v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2404.12377v1", "finalDownloadUrl": "https://arxiv.org/pdf/2404.12377", "sha256": "7faf7287c5920d0c16b1f55849a214d97b8355af621632176178860d07b237e0", "bytes": 2747356, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:37:33.902586+00:00", "publicationDate": "2024/04/18", "publicationDateSourceUrl": "https://arxiv.org/abs/2404.12377", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2404.12377", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2404.12377", "pages": 12, "sourceIdentityStatus": "title_match", "sourceTitle": "RoboDreamer: Learning Compositional World Models for Robot Imagination", "sourceVersion": "2404.12377v1", "sourceVersionDate": "2024/04/18", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.986124+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2404.12377" }, "originalAnalysis": { "id": "arxiv-2404.12377", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090134+00:00", "sourceUrl": "https://arxiv.org/html/2404.12377v1", "sectionsRead": [ "§2.2–3.3", "§4.1–4.2", "Tables 1–3", "§6", "Appendix A" ], "methodsZh": "将动作语言拆为动词和空间关系短语，组合各条件的视频扩散得分来生成未来图像计划；可再加入目标图或草图。执行时用相邻预测帧与当前状态训练逆动力学模型把视频转成动作，因而视频想象与控制是两个模块。", "experimentsZh": "以RT-1约7万真实记录、约500任务训练视频模型，抽取未见语言组合；约128样本由至少三人评估。机器人规划在RLBench仿真Panda上使用前视RGB与宏步动作，表3列六项任务，对照Image-BC、Hiveformer、UniPi。", "resultsZh": "作者报告未见指令生成视频的人评81.3%，HiP50.1%；加草图/目标图达到94.7%/95.8%。六项RLBench任务平均49.3%，Hiveformer44.2%，但堆块18.5%、取鞋10.5%，难任务仍较低。", "limitationsZh": "视频人评是想象计划合理性，不是实机成功率；真实RT-1数据也不构成新实体部署。论文明确单视角、真实新图像泛化不足和移动相机困难，且宏步设置弱化了低层连续控制。", "reproductionZh": "复现需保留语言组合拆分、数据划分、生成分辨率和逆动力学训练，区分纯语言与额外目标图条件。附录给出约100张V100、三阶段扩散及逆模型配置；未据RLBench默认硬件推断真实Panda实验。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "Table 3", "note": "六项RLBench均值49.3%；非实机。" }, { "section/page": "§4.1 / Appendix A.2", "note": "真实数据用于生成评估，至少3个人评、约128样本。" }, { "section/page": "§6", "note": "单视角、真实图像和移动相机限制。" } ], "corrections": { "robots": [], "robotNote": "Panda为RLBench仿真本体，未报告实体部署。" } }, "analysisVerifiedAt": "2026-10-04T13:50:35.090134+00:00" }, { "id": "arxiv-2404.05695", "title": "Humanoid-Gym: Reinforcement Learning for Humanoid Robot with Zero-Shot Sim2Real Transfer", "titleZh": "Humanoid-Gym：实现零样本仿真到现实迁移的人形机器人强化学习框架", "shortTitle": "Humanoid-Gym", "date": "2024-04-08", "datePrecision": "day", "year": 2024, "url": "https://arxiv.org/abs/2404.05695", "paperUrl": "https://arxiv.org/abs/2404.05695", "codeUrl": "https://github.com/roboterax/humanoid-gym", "summary": "用PPO及非对称观测训练人形行走，策略输入本体历史、速度命令和周期相位，输出关节PD目标。正弦参考、单双支撑掩模与速度/姿态/平滑奖励共同塑造步态；在Isaac Gym大规模随机化训练后先转MuJoCo校验。", "abstractZh": "用PPO及非对称观测训练人形行走，策略输入本体历史、速度命令和周期相位，输出关节PD目标。正弦参考、单双支撑掩模与速度/姿态/平滑奖励共同塑造步态；在Isaac Gym大规模随机化训练后先转MuJoCo校验。\n附录配置为8192并行环境、15帧策略历史和3帧特权历史，随机化传感噪声、延迟及动力学。用Robot Era XBot-S和XBot-L两尺寸机器人验证迁移；MuJoCo包含平地与不平地，并以真实腿摆正弦响应和相图校准模型。", "category": "人形机器人 / 训练框架", "tags": [ "PPO", "Isaac Gym", "MuJoCo", "Sim2Real" ], "directions": [ "人形机器人", "运动控制", "强化学习" ], "tier": "classic", "experimentType": "both", "robots": [ "Robot Era XBot-S", "Robot Era XBot-L" ], "robotNote": "", "codeStatus": "open", "status": "官方PPO训练、策略导出和sim-to-sim脚本公开。", "trainingNote": "仓库以XBot-L为主要示例，提供4096环境PPO训练、play导出和MuJoCo检查；依赖Isaac Gym Preview 4。", "contribution": "用PPO及非对称观测训练人形行走，策略输入本体历史、速度命令和周期相位，输出关节PD目标。正弦参考、单双支撑掩模与速度/姿态/平滑奖励共同塑造步态；在Isaac Gym大规模随机化训练后先转MuJoCo校验。", "whyUseful": "适合作为开源训练/迁移框架入口，需依附录恢复奖励、随机化与时延，逐级检查站立、关节响应及低速行走。实际复现需安全保护和每平台参数核对，本次未运行训练或硬件。", "limitations": "零样本是训练后无需真实策略微调，并不排除硬件测量和仿真校准。论文较短，步态奖励和机械参数仍依平台设计；不能把MuJoCo不平地演示直接当作同难度实机场景验证。", "evidence": [ { "url": "https://arxiv.org/abs/2404.05695", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://github.com/roboterax/humanoid-gym", "note": "官方README明确XBot-S/XBot-L、PPO脚本、导出及MuJoCo流程，区分未来功能。" }, { "url": "https://arxiv.org/pdf/2404.05695", "note": "§III; Appendix Tables II–IV：PPO、相位奖励、8192环境及历史帧配置。" }, { "url": "https://arxiv.org/pdf/2404.05695", "note": "§IV; Figures 3–7：两尺寸机器人、MuJoCo校准与定性迁移证据。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2024：人形Sim2Real研究形成更易上手的开源训练与验证工具链。", "projectUrl": null, "robotFilters": [ "Robot Era XBot-S", "Robot Era XBot-L" ], "original": { "id": "arxiv-2404.05695", "originalSourceUrl": "https://arxiv.org/pdf/2404.05695v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2404.05695v2", "finalDownloadUrl": "https://arxiv.org/pdf/2404.05695", "sha256": "f5a95d5797dc0b788087fc454fef0923f4d5055a14d69645902c0fcbc96e097e", "bytes": 3893938, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:37:39.861664+00:00", "publicationDate": "2024/04/08", "publicationDateSourceUrl": "https://arxiv.org/abs/2404.05695", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2404.05695", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2404.05695", "pages": 6, "sourceIdentityStatus": "title_match", "sourceTitle": "Humanoid-Gym: Reinforcement Learning for Humanoid Robot with Zero-Shot Sim2Real Transfer", "sourceVersion": "2404.05695v2", "sourceVersionDate": "2024/05/18", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.989559+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2404.05695" }, "originalAnalysis": { "id": "arxiv-2404.05695", "sourceUrl": "https://arxiv.org/pdf/2404.05695", "sectionsRead": [ "§III", "§IV", "Appendix Tables II–IV", "Figures 3–7" ], "methodsZh": "用PPO及非对称观测训练人形行走，策略输入本体历史、速度命令和周期相位，输出关节PD目标。正弦参考、单双支撑掩模与速度/姿态/平滑奖励共同塑造步态；在Isaac Gym大规模随机化训练后先转MuJoCo校验。", "experimentsZh": "附录配置为8192并行环境、15帧策略历史和3帧特权历史，随机化传感噪声、延迟及动力学。用Robot Era XBot-S和XBot-L两尺寸机器人验证迁移；MuJoCo包含平地与不平地，并以真实腿摆正弦响应和相图校准模型。", "resultsZh": "作者展示两平台零样本sim-to-real行走，并报告校准后MuJoCo腿部轨迹与真实系统较接近。结果主要是视频及轨迹对照，没有系统给出多种子学习曲线、跨地形成功率或统计置信区间，不能补造稳定性分数。", "limitationsZh": "零样本是训练后无需真实策略微调，并不排除硬件测量和仿真校准。论文较短，步态奖励和机械参数仍依平台设计；不能把MuJoCo不平地演示直接当作同难度实机场景验证。", "reproductionZh": "适合作为开源训练/迁移框架入口，需依附录恢复奖励、随机化与时延，逐级检查站立、关节响应及低速行走。实际复现需安全保护和每平台参数核对，本次未运行训练或硬件。", "experimentType": "both", "robots": [ "Robot Era XBot-S", "Robot Era XBot-L" ], "evidenceNotes": [ { "section/page": "§III; Appendix Tables II–IV", "note": "PPO、相位奖励、8192环境及历史帧配置。" }, { "section/page": "§IV; Figures 3–7", "note": "两尺寸机器人、MuJoCo校准与定性迁移证据。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "附录配置为8192并行环境、15帧策略历史和3帧特权历史，随机化传感噪声、延迟及动力学。用Robot Era XBot-S和XBot-L两尺寸机器人验证迁移；MuJoCo包含平地与不平地，并以真实腿摆正弦响应和相图校准模型。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2403.12945", "title": "DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset", "titleZh": "DROID：大规模自然场景机器人操作数据集", "shortTitle": "DROID", "date": "2024-03-19", "year": 2024, "url": "https://arxiv.org/abs/2403.12945", "codeUrl": "https://github.com/droid-dataset/droid_policy_learning", "category": "数据集与基准", "tags": [ "真实数据", "示范采集", "场景多样性", "Franka" ], "tier": "foundation", "summary": "DROID以一致且可推行的数据采集硬件分布到18处，由50人收集不同建筑、任务、视角与场景；Panda加Robotiq 2F-85、三路ZED相机及Quest遥操作，15Hz同步动作。它主要贡献真实数据分布与采…", "abstractZh": "DROID以一致且可推行的数据采集硬件分布到18处，由50人收集不同建筑、任务、视角与场景；Panda加Robotiq 2F-85、三路ZED相机及Quest遥操作，15Hz同步动作。它主要贡献真实数据分布与采集流程，并用标准语言条件扩散策略检验价值。\n核心集7.6万成功轨迹、564场景、52建筑，另发布约1.6万失败轨迹但不计核心规模。六个实验室/办公室/家庭任务各收50–150条目标示范，与DROID或OXE按50/50混批；每条件方法10次，含未见对象、干扰物等。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "real", "robots": [ "Franka Emika Panda", "Robotiq 2F-85" ], "codeStatus": "open", "status": "数据、平台采集代码与策略训练/评测代码公开。", "trainingNote": "所链接策略学习仓库为 MIT；采集硬件仓库许可应另核。数据集许可与代码分开，不将可下载自动等同于任意用途授权。", "whyUseful": "先复现50/50混批、16步预测/8步执行及固定目标数据，避免测试场景泄漏；比较等轨迹量的场景多样性。失败集与成功集应分别标识，下载和存储预算按视频规模估计，不能把全部数据都称成功示范。", "contribution": "DROID以一致且可推行的数据采集硬件分布到18处，由50人收集不同建筑、任务、视角与场景；Panda加Robotiq 2F-85、三路ZED相机及Quest遥操作，15Hz同步动作。它主要贡献真实数据分布与采集流程，并用标准语言条件扩散策略检验价值。", "limitations": "单一Panda形态，不能证明跨机器人迁移；每条件仅10次且同域示范仍必需，整体均值不能遮蔽特定对象负迁移。相机外参、场景标签和后处理质量影响下游，多样性不是无条件保证泛化。", "caveats": "带轮升降桌用于搬迁采集场地，不代表策略学习自主移动导航。", "evidence": [ { "url": "https://arxiv.org/html/2403.12945v1", "note": "首发日期、第 III 节硬件和数据采集、第 V 节实机评测。" }, { "url": "https://droid-dataset.github.io/", "note": "官方数据及硬件入口。" }, { "url": "https://github.com/droid-dataset/droid_policy_learning", "note": "实际训练、评测代码与 MIT 许可证。" }, { "url": "https://arxiv.org/pdf/2403.12945", "note": "III; V-A：采集硬件、成功/失败规模及联合训练协议。" }, { "url": "https://arxiv.org/pdf/2403.12945", "note": "p.9 Fig.8（已渲染）; V-C：75/53均值、个别负迁移及场景数量消融。" }, { "url": "https://arxiv.org/pdf/2403.12945v2", "note": "III; V-A：采集硬件、成功/失败规模及联合训练协议。" }, { "url": "https://arxiv.org/pdf/2403.12945v2", "note": "p.9 Fig.8（已渲染）; V-C：75/53均值、个别负迁移及场景数量消融。" } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "操作与抓取", "数据集与基准", "模仿学习" ], "robotFilters": [ "Franka Panda", "Robotiq 2F-85" ], "verificationNote": "已核原论文硬件与实机实验；代码链接指向有许可证的策略实现。", "original": { "id": "arxiv-2403.12945", "originalSourceUrl": "https://arxiv.org/pdf/2403.12945v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2403.12945v2", "finalDownloadUrl": "https://arxiv.org/pdf/2403.12945", "sha256": "9e40a9c934ce78ea315d2c05593d60f9e18edf6dcd3785cd66439518965c3b07", "bytes": 9707714, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:37:46.824880+00:00", "publicationDate": "2024/03/19", "publicationDateSourceUrl": "https://arxiv.org/abs/2403.12945", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2403.12945", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2403.12945", "pages": 26, "sourceIdentityStatus": "title_match", "sourceTitle": "DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset", "sourceVersion": "2403.12945v2", "sourceVersionDate": "2025/04/22", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:14.996561+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2403.12945" }, "originalAnalysis": { "id": "arxiv-2403.12945", "sourceUrl": "https://arxiv.org/pdf/2403.12945v2", "sectionsRead": [ "III-A–B", "IV", "V-A–C", "VI", "Fig.8（渲染核对）" ], "methodsZh": "DROID以一致且可推行的数据采集硬件分布到18处，由50人收集不同建筑、任务、视角与场景；Panda加Robotiq 2F-85、三路ZED相机及Quest遥操作，15Hz同步动作。它主要贡献真实数据分布与采集流程，并用标准语言条件扩散策略检验价值。", "experimentsZh": "核心集7.6万成功轨迹、564场景、52建筑，另发布约1.6万失败轨迹但不计核心规模。六个实验室/办公室/家庭任务各收50–150条目标示范，与DROID或OXE按50/50混批；每条件方法10次，含未见对象、干扰物等。", "resultsZh": "图8分布内平均成功75%，无联合训练53%、OXE43%；分布外53%，对17%、36%。并非每个条件都胜出，未见薯片OXE为60%、DROID30%。等约7000条数据但更多场景的子集更利于分布外表现。", "limitationsZh": "单一Panda形态，不能证明跨机器人迁移；每条件仅10次且同域示范仍必需，整体均值不能遮蔽特定对象负迁移。相机外参、场景标签和后处理质量影响下游，多样性不是无条件保证泛化。", "reproductionZh": "先复现50/50混批、16步预测/8步执行及固定目标数据，避免测试场景泄漏；比较等轨迹量的场景多样性。失败集与成功集应分别标识，下载和存储预算按视频规模估计，不能把全部数据都称成功示范。", "experimentType": "real", "robots": [ "Franka Emika Panda", "Robotiq 2F-85" ], "corrections": [ "7.6万指成功核心集，约1.6万失败另发布；未见薯片条件DROID并非最优。" ], "evidenceNotes": [ { "note": "采集硬件、成功/失败规模及联合训练协议。", "section": "III; V-A" }, { "note": "75/53均值、个别负迁移及场景数量消融。", "section": "p.9 Fig.8（已渲染）; V-C" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2403.12945v2", "originalSha256": "9e40a9c934ce78ea315d2c05593d60f9e18edf6dcd3785cd66439518965c3b07" }, "experimentNote": "核心集7.6万成功轨迹、564场景、52建筑，另发布约1.6万失败轨迹但不计核心规模。六个实验室/办公室/家庭任务各收50–150条目标示范，与DROID或OXE按50/50混批；每条件方法10次，含未见对象、干扰物等。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2403.09227", "title": "BEHAVIOR-1K: A Human-Centered, Embodied AI Benchmark with 1,000 Everyday Activities and Realistic Simulation", "date": "2024-03-14", "url": "https://arxiv.org/abs/2403.09227v1", "titleZh": "BEHAVIOR-1K：以人类需求为中心的一千项日常活动仿真基准", "abstractZh": "把人类家务需求表示为BDDL语义目标，配合场景/物体资产和OmniGibson的刚体、软体、液体仿真，支持长时序活动研究。\n从1000项活动中选收垃圾、收纳装饰和擦桌等少数任务评估；在模拟公寓与真实TIAGo上比较收垃圾的策略执行。", "summary": "把人类家务需求表示为BDDL语义目标，配合场景/物体资产和OmniGibson的刚体、软体、液体仿真，支持长时序活动研究。", "experimentType": "both", "robots": [ "PAL Robotics TIAGo双臂移动操作平台" ], "tags": [ "数据集与基准", "操作与抓取", "导航与建图" ], "limitations": "不能把1000个任务定义理解为已学会1000项技能；部分仿真使用辅助抓取与完美定位，原版不包括与人互动的活动。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "foundation", "tierNote": "编辑分类：以日常活动定义、语义目标与仿真资产为核心的基准基础设施；不表示所有任务已被解决。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2403.09227v1", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2403.09227v1", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/StanfordVL/BEHAVIOR-1K", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/StanfordVL/BEHAVIOR-1K", "category": "数据集与基准 / 操作与抓取", "year": 2024, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "提供模拟器与训练/评估基础设施；具体基线要按对应发布版本核对。 有资产与数据模块，下载和使用条件依各发布页。 不是单一预训练策略；本次未核实特定挑战基线权重。 机器人列表是当前模拟器支持情况，不证明每种机器人都复现过原始论文全部任务；资产与平台版本必须固定。 本次未运行训练、复现实验或实机控制。", "directions": [ "数据集与基准", "操作与抓取", "导航与建图" ], "robotFilters": [ "PAL Robotics TIAGo双臂移动操作平台" ], "original": { "id": "arxiv-2403.09227", "originalSourceUrl": "https://arxiv.org/pdf/2403.09227v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2403.09227v1", "finalDownloadUrl": "https://arxiv.org/pdf/2403.09227v1", "sha256": "ab4033f3e8a8187ad2cfdf11f1979d33996ecce1e81fa3d62af33a92bbb17a2b", "bytes": 14233640, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:15.230868+00:00", "publicationDate": "2024/03/14", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2403.09227v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2403.09227v1", "pages": 43, "sourceIdentityStatus": "title_match", "sourceTitle": "BEHAVIOR-1K: A Human-Centered, Embodied AI Benchmark with 1,000 Everyday Activities and Realistic Simulation", "sourceVersion": "2403.09227v1", "sourceVersionDate": "2024/03/14", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2403.09227v1", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2403.09227" }, "originalAnalysis": { "id": "arxiv-2403.09227", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2403.09227v1", "sourceVersion": "2403.09227v1", "sectionsRead": [ "PDF第5页选段（提取文本L1029–1046）", "PDF第7页选段（提取文本L1156–1169）", "PDF第8页选段（提取文本L1315–1332）", "PDF第9页选段（提取文本L1369–1386）", "PDF第6–7页选段（提取文本L1124–1145）", "PDF第8–9页选段（提取文本L1332–1365）", "PDF第9页选段（提取文本L1385–1404）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "把人类家务需求表示为BDDL语义目标，配合场景/物体资产和OmniGibson的刚体、软体、液体仿真，支持长时序活动研究。", "experimentsZh": "从1000项活动中选收垃圾、收纳装饰和擦桌等少数任务评估；在模拟公寓与真实TIAGo上比较收垃圾的策略执行。", "resultsZh": "真实迁移中学习策略26次测试均未完成；人工选择原语约22%成功。感知、抓取和导航累积误差揭示显著仿真到现实差距。", "limitationsZh": "不能把1000个任务定义理解为已学会1000项技能；部分仿真使用辅助抓取与完美定位，原版不包括与人互动的活动。", "reproductionZh": "保留BDDL任务版本、资产条款、原语和抓取简化设置；分开报告理想策略、学习策略、仿真及实机结果。", "experimentType": "both", "robots": [ "PAL Robotics TIAGo双臂移动操作平台" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "把人类家务需求表示为BDDL语义目标，配合场景/物体资产和OmniGibson的刚体、软体、液体仿真，支持长时序活动研究。" }, { "section": "原文实验与结果相关选段", "note": "从1000项活动中选收垃圾、收纳装饰和擦桌等少数任务评估；在模拟公寓与真实TIAGo上比较收垃圾的策略执行。 真实迁移中学习策略26次测试均未完成；人工选择原语约22%成功。感知、抓取和导航累积误差揭示显著仿真到现实差距。" }, { "section": "局限与复现条件", "note": "不能把1000个任务定义理解为已学会1000项技能；部分仿真使用辅助抓取与完美定位，原版不包括与人互动的活动。" } ], "originalSha256": "ab4033f3e8a8187ad2cfdf11f1979d33996ecce1e81fa3d62af33a92bbb17a2b", "readingScopeSha256": "e91c6432498da83aa76e0828209842d7d55c22e41bb03d4a04d2cdecbcc7e74e", "textExtractionSha256": "c5bc23bd0b6acf0024d35ab1d7558200edfb9b7194833792caabf75167ffa580" }, "experimentNote": "从1000项活动中选收垃圾、收纳装饰和擦桌等少数任务评估；在模拟公寓与真实TIAGo上比较收垃圾的策略执行。", "contribution": "把人类家务需求表示为BDDL语义目标，配合场景/物体资产和OmniGibson的刚体、软体、液体仿真，支持长时序活动研究。", "whyUseful": "保留BDDL任务版本、资产条款、原语和抓取简化设置；分开报告理想策略、学习策略、仿真及实机结果。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "根仓库 MIT；Isaac Sim、物体/场景资产和数据集存在独立条款，不能将整个资产库视为 MIT。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2403.07788", "title": "DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation", "date": "2024-03-12", "url": "https://arxiv.org/abs/2403.07788v2", "titleZh": "DexCap：便携式灵巧操作动捕采集系统", "abstractZh": "便携SLAM与电磁动捕记录手腕、手指和三维场景；以指尖逆运动学重定向到LEAP手，点云扩散策略学习动作，并可加入人在环纠正。\n双Franka机械臂各装LEAP Hand，在海绵抓取、收球、擦盘、包装、剪刀及泡茶六类任务评价，区分子任务与完整任务成功。", "summary": "便携SLAM与电磁动捕记录手腕、手指和三维场景；以指尖逆运动学重定向到LEAP手，点云扩散策略学习动作，并可加入人在环纠正。", "experimentType": "real", "robots": [ "Franka Emika机械臂（原文未细分型号）", "LEAP Hand" ], "tags": [ "灵巧手", "数据集与基准", "模仿学习" ], "limitations": "原版单次采集续航最多约40分钟；人体与机械手尺寸、手指数不同，指尖匹配不保证接触力一致，系统未采集力觉。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的便携多指示教系统研究，关注动作重定向与接触信息缺失。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2403.07788v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2403.07788v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/j96w/DexCap", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/j96w/DexCap", "category": "灵巧手 / 数据集与基准", "year": 2024, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "提供训练脚本和配置，输出双手双臂动作序列。 Hugging Face 提供原始和处理后数据链接。 本次未核实公开预训练策略检查点。 代码开放不代表动捕商业软件开放；需处理 SLAM 漂移、坐标对齐及手部运动学差异。 本次未运行训练、复现实验或实机控制。", "directions": [ "灵巧手", "数据集与基准", "模仿学习" ], "robotFilters": [ "Franka Emika机械臂（原文未细分型号）", "LEAP Hand" ], "original": { "id": "arxiv-2403.07788", "originalSourceUrl": "https://arxiv.org/pdf/2403.07788v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2403.07788v2", "finalDownloadUrl": "https://arxiv.org/pdf/2403.07788v2", "sha256": "ad87015fd70061e5d0b6d4b21c22411c8071328766892a55f133b652d16de87a", "bytes": 23271528, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:20.780871+00:00", "publicationDate": "2024/03/12", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2403.07788v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2403.07788v2", "pages": 20, "sourceIdentityStatus": "title_match", "sourceTitle": "DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation", "sourceVersion": "2403.07788v2", "sourceVersionDate": "2024/07/04", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2403.07788v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2403.07788" }, "originalAnalysis": { "id": "arxiv-2403.07788", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2403.07788v2", "sourceVersion": "2403.07788v2", "sectionsRead": [ "PDF第3页选段（提取文本L231–248）", "PDF第4–5页选段（提取文本L353–366）", "PDF第6页选段（提取文本L543–556）", "PDF第7页选段（提取文本L631–648）", "PDF第11页选段（提取文本L1003–1020）", "PDF第5–6页选段（提取文本L463–506）", "PDF第8–9页选段（提取文本L713–739）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "便携SLAM与电磁动捕记录手腕、手指和三维场景；以指尖逆运动学重定向到LEAP手，点云扩散策略学习动作，并可加入人在环纠正。", "experimentsZh": "双Franka机械臂各装LEAP Hand，在海绵抓取、收球、擦盘、包装、剪刀及泡茶六类任务评价，区分子任务与完整任务成功。", "resultsZh": "动捕相对纯视觉手估计更耐遮挡，论文报告采集吞吐约为遥操作三倍；精密工具使用仍受益于实机纠正。", "limitationsZh": "原版单次采集续航最多约40分钟；人体与机械手尺寸、手指数不同，指尖匹配不保证接触力一致，系统未采集力觉。", "reproductionZh": "复现世界坐标对齐、末端点云处理和重定向比例；必须披露纠正数据量，不能把人类纠正后的结果算成纯人类视频零样本。", "experimentType": "real", "robots": [ "Franka Emika机械臂（原文未细分型号）", "LEAP Hand" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "便携SLAM与电磁动捕记录手腕、手指和三维场景；以指尖逆运动学重定向到LEAP手，点云扩散策略学习动作，并可加入人在环纠正。" }, { "section": "原文实验与结果相关选段", "note": "双Franka机械臂各装LEAP Hand，在海绵抓取、收球、擦盘、包装、剪刀及泡茶六类任务评价，区分子任务与完整任务成功。 动捕相对纯视觉手估计更耐遮挡，论文报告采集吞吐约为遥操作三倍；精密工具使用仍受益于实机纠正。" }, { "section": "局限与复现条件", "note": "原版单次采集续航最多约40分钟；人体与机械手尺寸、手指数不同，指尖匹配不保证接触力一致，系统未采集力觉。" } ], "originalSha256": "ad87015fd70061e5d0b6d4b21c22411c8071328766892a55f133b652d16de87a", "readingScopeSha256": "8b66e35e73c4e4c0013e583068a7e65a27e397b46a6bbf6b7a588a4fbd6b8865", "textExtractionSha256": "b8c83462815bcca1e0a577ee6764e3f651ca41c2d33bf4618bd6000770df6586" }, "experimentNote": "双Franka机械臂各装LEAP Hand，在海绵抓取、收球、擦盘、包装、剪刀及泡茶六类任务评价，区分子任务与完整任务成功。", "contribution": "便携SLAM与电磁动捕记录手腕、手指和三维场景；以指尖逆运动学重定向到LEAP手，点云扩散策略学习动作，并可加入人在环纠正。", "whyUseful": "复现世界坐标对齐、末端点云处理和重定向比例；必须披露纠正数据量，不能把人类纠正后的结果算成纯人类视频零样本。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "仓库 MIT；Rokoko Studio 等外部商业软件、控制器、数据权利需另核对。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2403.04436", "title": "Learning Human-to-Humanoid Real-Time Whole-Body Teleoperation", "titleZh": "学习从人到人形机器人的实时全身遥操作", "shortTitle": "H2O", "date": "2024-03-07", "datePrecision": "day", "year": 2024, "url": "https://arxiv.org/abs/2403.04436", "paperUrl": "https://arxiv.org/abs/2403.04436", "codeUrl": "https://github.com/LeCAR-Lab/human2humanoid", "summary": "H2O先调整SMPL人体比例并以12个对应关节点重定向AMASS动作，再训练可见完整刚体状态的特权策略，删除其无法跟踪的序列。最终控制器只用可获得的本体状态及八个参考关键点，输出19维关节目标，经PD控制执行；…", "abstractZh": "H2O先调整SMPL人体比例并以12个对应关节点重定向AMASS动作，再训练可见完整刚体状态的特权策略，删除其无法跟踪的序列。最终控制器只用可获得的本体状态及八个参考关键点，输出19维关节目标，经PD控制执行；摩擦、质量、增益、延迟等随机化用于实机迁移。\n约1万条重定向动作筛为8500条可行动作，仿真仍在未经清理的1万条上比较全系统、无清理和简化目标状态。真实Unitree H1通过30Hz HybrIK读取RGB人体姿态，演示行走、踢球、推车和后跳，另做外力扰动测试。", "category": "人形机器人 / 全身遥操作", "tags": [ "人类动作", "运动重定向", "实时遥操作", "Sim2Real" ], "directions": [ "人形机器人", "运动控制", "操作与抓取" ], "tier": "classic", "experimentType": "both", "robots": [ "Unitree H1" ], "robotNote": "", "codeStatus": "open", "status": "作者H2O与OmniH2O共享训练仓库公开，许可限制非商业用途。", "trainingNote": "需要动作重定向、sim-to-data筛选与域随机化训练；仓库基于Isaac Gym、legged_gym和rsl_rl。", "contribution": "H2O先调整SMPL人体比例并以12个对应关节点重定向AMASS动作，再训练可见完整刚体状态的特权策略，删除其无法跟踪的序列。最终控制器只用可获得的本体状态及八个参考关键点，输出19维关节目标，经PD控制执行；摩擦、质量、增益、延迟等随机化用于实机迁移。", "whyUseful": "复现需AMASS/SMPL处理、H1模型、两阶段训练、精确PD和随机化配置，还须提供根部速度估计。应单独测端到端延迟、失衡率和未经筛选的新动作，不能仅以展示视频验证鲁棒性。", "limitations": "所谓“仅RGB相机”针对操作者动作输入；机器人根部线速度实际由50Hz动捕系统估计。可行动作筛选受特权策略能力限制，视觉延迟、形态差异及随机化强度仍影响精度，不能理解为自主任务规划。", "evidence": [ { "url": "https://arxiv.org/abs/2403.04436", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/html/2403.04436v1", "note": "正文明确Unitree H1；实机设置写明RGB相机估计人姿态、动捕估计机器人线速度。" }, { "url": "https://human2humanoid.com/", "note": "作者项目展示sim-to-data、重定向与部署流程。" }, { "url": "https://github.com/LeCAR-Lab/human2humanoid", "note": "官方实现与CC BY-NC 4.0许可声明。" }, { "url": "https://arxiv.org/pdf/2403.04436v1", "note": "IV-A/B：13k AMASS sequences →10k retargeted →~8.5k cleaned; privileged controller filters failures." }, { "url": "https://arxiv.org/pdf/2403.04436v1", "note": "VI-A 表III：72.5%, 67.9%, 53.2% are simulation imitation success on 10k sequences; >0.5m body distance fails." }, { "url": "https://arxiv.org/pdf/2403.04436v1", "note": "VI-B Deployment Details：Unitree H1; 30Hz pose estimation; external 50Hz MoCap for linear velocity; onboard sensors 200Hz." }, { "url": "https://arxiv.org/pdf/2403.04436v1", "note": "VII：Pose latency, embodiment feasibility and randomization/imitation trade-offs explicitly discussed." } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2024：大规模人类动作清理与仿真RL连接到全尺寸人形实时跟随。", "projectUrl": null, "robotFilters": [ "Unitree H1" ], "original": { "id": "arxiv-2403.04436", "originalSourceUrl": "https://arxiv.org/pdf/2403.04436v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2403.04436v1", "finalDownloadUrl": "https://arxiv.org/pdf/2403.04436", "sha256": "78d380da6e28949275137a7dd46922116a81efdf417a02aa81c6f40c5bcade22", "bytes": 13325640, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:37:47.387849+00:00", "publicationDate": "2024/03/07", "publicationDateSourceUrl": "https://arxiv.org/abs/2403.04436", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2403.04436", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2403.04436", "pages": 10, "sourceIdentityStatus": "title_match", "sourceTitle": "Learning Human-to-Humanoid Real-Time Whole-Body Teleoperation", "sourceVersion": "2403.04436v1", "sourceVersionDate": "2024/03/07", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.008460+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2403.04436" }, "originalAnalysis": { "id": "arxiv-2403.04436", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739389+00:00", "sourceUrl": "https://arxiv.org/pdf/2403.04436v1", "sectionsRead": [ "IV-A/B 重定向与sim-to-data", "V-A–D 状态、奖励、随机化和终止", "VI 表III/IV及实机部署", "VII 讨论与限制" ], "methodsZh": "H2O先调整SMPL人体比例并以12个对应关节点重定向AMASS动作，再训练可见完整刚体状态的特权策略，删除其无法跟踪的序列。最终控制器只用可获得的本体状态及八个参考关键点，输出19维关节目标，经PD控制执行；摩擦、质量、增益、延迟等随机化用于实机迁移。", "experimentsZh": "约1万条重定向动作筛为8500条可行动作，仿真仍在未经清理的1万条上比较全系统、无清理和简化目标状态。真实Unitree H1通过30Hz HybrIK读取RGB人体姿态，演示行走、踢球、推车和后跳，另做外力扰动测试。", "resultsZh": "作者表III报告仿真跟踪成功率72.5%，不清理为67.9%，简化状态为53.2%；成功定义包括全程平均身体偏差不超过0.5米。特权策略85.5%不能直接部署。真实实验展示多样动态行为，没有同等规模的实机成功率统计。", "limitationsZh": "所谓“仅RGB相机”针对操作者动作输入；机器人根部线速度实际由50Hz动捕系统估计。可行动作筛选受特权策略能力限制，视觉延迟、形态差异及随机化强度仍影响精度，不能理解为自主任务规划。", "reproductionZh": "复现需AMASS/SMPL处理、H1模型、两阶段训练、精确PD和随机化配置，还须提供根部速度估计。应单独测端到端延迟、失衡率和未经筛选的新动作，不能仅以展示视频验证鲁棒性。", "evidenceNotes": [ { "section/page": "IV-A/B", "note": "13k AMASS sequences →10k retargeted →~8.5k cleaned; privileged controller filters failures." }, { "section/page": "VI-A 表III", "note": "72.5%, 67.9%, 53.2% are simulation imitation success on 10k sequences; >0.5m body distance fails." }, { "section/page": "VI-B Deployment Details", "note": "Unitree H1; 30Hz pose estimation; external 50Hz MoCap for linear velocity; onboard sensors 200Hz." }, { "section/page": "VII", "note": "Pose latency, embodiment feasibility and randomization/imitation trade-offs explicitly discussed." } ], "experimentType": "both", "robots": [ "Unitree H1" ], "corrections": [ "“仅RGB”不意味着整个部署无外部动捕；原文用MoCap估计机器人线速度。" ], "sourceVersion": "2403.04436v1", "originalSha256": "78d380da6e28949275137a7dd46922116a81efdf417a02aa81c6f40c5bcade22" }, "experimentNote": "约1万条重定向动作筛为8500条可行动作，仿真仍在未经清理的1万条上比较全系统、无清理和简化目标状态。真实Unitree H1通过30Hz HybrIK读取RGB人体姿态，演示行走、踢球、推车和后跳，另做外力扰动测试。", "analysisVerifiedAt": "2026-10-04T13:47:58.739389+00:00" }, { "id": "arxiv-2403.03954", "title": "3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations", "date": "2024-03-06", "url": "https://arxiv.org/abs/2403.03954v7", "titleZh": "三维扩散策略：用简单三维表示实现可泛化视觉运动学习", "abstractZh": "DP3把稀疏三维点云压缩为紧凑视觉表示，与机器人状态共同条件化扩散动作序列；简单的三维编码避免复杂体素或大规模点云骨干。\n72项仿真任务及四项真实操作任务；真实系统包括Franka机械臂、Allegro手或夹爪，RealSense L515提供视觉。", "summary": "DP3把稀疏三维点云压缩为紧凑视觉表示，与机器人状态共同条件化扩散动作序列；简单的三维编码避免复杂体素或大规模点云骨干。", "experimentType": "both", "robots": [ "Franka机械臂（原文未细分型号）", "Allegro Hand" ], "tags": [ "模仿学习", "操作与抓取", "灵巧手" ], "limitations": "最优三维控制表示尚未确定，未深入测试极长时序任务；深度质量及工作空间裁剪仍影响部署。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的点云扩散策略研究，适合分析视觉表示与泛化的关系。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2403.03954v7", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2403.03954v7", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/YanjieZe/3D-Diffusion-Policy", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/YanjieZe/3D-Diffusion-Policy", "category": "模仿学习 / 操作与抓取", "year": 2024, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "配置与训练入口已发布；真实部署参考另外的 iDP3 项目。 仿真环境与专家策略用于生成数据；提供真实机器人数据链接。 有专家策略；本次未核实覆盖全部任务的预训练 DP3 权重。 README 同时列出后续项目，不能把后续结果当成原始 DP3；真实部署需要相机标定、点云裁剪和机器人控制器。 本次未运行训练、复现实验或实机控制。", "directions": [ "模仿学习", "操作与抓取", "灵巧手" ], "robotFilters": [ "Franka机械臂（原文未细分型号）", "Allegro Hand" ], "original": { "id": "arxiv-2403.03954", "originalSourceUrl": "https://arxiv.org/pdf/2403.03954v7", "versionedOriginalUrl": "https://arxiv.org/pdf/2403.03954v7", "finalDownloadUrl": "https://arxiv.org/pdf/2403.03954v7", "sha256": "916ecb6011e5e311daa9a8ca53d0ec2d8f44f7d2034d9ba368de7102737d9682", "bytes": 14149153, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T14:59:42.736715+00:00", "publicationDate": "2024/03/06", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2403.03954v7", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2403.03954v7", "pages": 17, "sourceIdentityStatus": "title_match", "sourceTitle": "3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations", "sourceVersion": "2403.03954v7", "sourceVersionDate": "2024/09/27", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2403.03954v7", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2403.03954" }, "originalAnalysis": { "id": "arxiv-2403.03954", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2403.03954v7", "sourceVersion": "2403.03954v7", "sectionsRead": [ "PDF第6页选段（提取文本L1032–1045）", "PDF第7页选段（提取文本L1605–1618）", "PDF第8页选段（提取文本L1777–1794）", "PDF第9–10页选段（提取文本L2072–2089）", "PDF第10页选段（提取文本L2194–2211）", "PDF第4页选段（提取文本L450–486）", "PDF第10页选段（提取文本L2194–2200）", "PDF第3页选段（提取文本L183–225）", "PDF第4–5页选段（提取文本L589–612）", "PDF第7–8页选段（提取文本L1740–1776）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "DP3把稀疏三维点云压缩为紧凑视觉表示，与机器人状态共同条件化扩散动作序列；简单的三维编码避免复杂体素或大规模点云骨干。", "experimentsZh": "72项仿真任务及四项真实操作任务；真实系统包括Franka机械臂、Allegro手或夹爪，RealSense L515提供视觉。", "resultsZh": "实验显示三维表示可改善学习效率和环境变化下的泛化；四项真实任务各评估10次，不能把小样本结果视作部署安全证明。", "limitationsZh": "最优三维控制表示尚未确定，未深入测试极长时序任务；深度质量及工作空间裁剪仍影响部署。", "reproductionZh": "复现点云裁剪、采样、相机标定及动作归一化，分别记录仿真种子和真实试次数；不沿用新仓库版本的结果覆盖论文版本。", "experimentType": "both", "robots": [ "Franka机械臂（原文未细分型号）", "Allegro Hand" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "DP3把稀疏三维点云压缩为紧凑视觉表示，与机器人状态共同条件化扩散动作序列；简单的三维编码避免复杂体素或大规模点云骨干。" }, { "section": "原文实验与结果相关选段", "note": "72项仿真任务及四项真实操作任务；真实系统包括Franka机械臂、Allegro手或夹爪，RealSense L515提供视觉。 实验显示三维表示可改善学习效率和环境变化下的泛化；四项真实任务各评估10次，不能把小样本结果视作部署安全证明。" }, { "section": "局限与复现条件", "note": "最优三维控制表示尚未确定，未深入测试极长时序任务；深度质量及工作空间裁剪仍影响部署。" } ], "originalSha256": "916ecb6011e5e311daa9a8ca53d0ec2d8f44f7d2034d9ba368de7102737d9682", "readingScopeSha256": "b5b87dbd250d69f2fe9e30bea4333f018347e14bf7e0c490bfc014773b6cd354", "textExtractionSha256": "62a910e2ed0c4df445d32f208bea076a824113db39ebd74a6aa852f9972dd054" }, "experimentNote": "72项仿真任务及四项真实操作任务；真实系统包括Franka机械臂、Allegro手或夹爪，RealSense L515提供视觉。", "contribution": "DP3把稀疏三维点云压缩为紧凑视觉表示，与机器人状态共同条件化扩散动作序列；简单的三维编码避免复杂体素或大规模点云骨干。", "whyUseful": "复现点云裁剪、采样、相机标定及动作归一化，分别记录仿真种子和真实试次数；不沿用新仓库版本的结果覆盖论文版本。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "核心代码 MIT；仿真资产、专家策略和外部控制软件应核对各自条款。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2402.10329", "title": "Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots", "date": "2024-02-15", "url": "https://arxiv.org/abs/2402.10329v3", "titleZh": "通用操作接口：无需把机器人带到现场的野外示教", "abstractZh": "UMI用带鱼眼相机、侧镜和IMU的手持夹爪记录示教，恢复相对末端轨迹；扩散策略配合观测同步与执行延迟匹配，减少人类采集与机器人部署的接口差异。\n四类真实任务覆盖单臂、双臂、动态及长时序操作；区分同环境能力测试与异环境泛化，并消融视野、动作表示和延迟处理。", "summary": "UMI用带鱼眼相机、侧镜和IMU的手持夹爪记录示教，恢复相对末端轨迹；扩散策略配合观测同步与执行延迟匹配，减少人类采集与机器人部署的接口差异。", "experimentType": "real", "robots": [ "Universal Robots UR5", "Franka Emika FR2" ], "tags": [ "操作与抓取", "模仿学习", "数据集与基准" ], "limitations": "采集时不知道下游运动学限制，仍需过滤不可达动作；视觉SLAM依赖场景纹理；手持夹爪比徒手示教笨重。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的机器人外示教与策略接口研究，适合研究数据采集和跨硬件部署条件。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2402.10329v3", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2402.10329v3", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/real-stanford/universal_manipulation_interface", "note": "已核验官方代码与文档；未运行复现；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/real-stanford/universal_manipulation_interface", "category": "操作与抓取 / 模仿学习", "year": 2024, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "README已核实SLAM处理、数据集生成及单/多GPU扩散策略训练命令；采集硬件、相机与机器人控制需单独配置。 本次未运行训练、复现实验或实机控制。", "directions": [ "操作与抓取", "模仿学习", "数据集与基准" ], "robotFilters": [ "Universal Robots UR5", "Franka Emika FR2" ], "original": { "id": "arxiv-2402.10329", "originalSourceUrl": "https://arxiv.org/pdf/2402.10329v3", "versionedOriginalUrl": "https://arxiv.org/pdf/2402.10329v3", "finalDownloadUrl": "https://arxiv.org/pdf/2402.10329v3", "sha256": "a9452cbaf9f5d2a77d5f8b631632fa7f45ae4bbc32400be1cc939b91a36e1609", "bytes": 7570839, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T14:59:31.229722+00:00", "publicationDate": "2024/02/15", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2402.10329v3", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2402.10329v3", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots", "sourceVersion": "2402.10329v3", "sourceVersionDate": "2024/03/06", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2402.10329v3", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2402.10329" }, "originalAnalysis": { "id": "arxiv-2402.10329", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2402.10329v3", "sourceVersion": "2402.10329v3", "sectionsRead": [ "PDF第6–7页选段（提取文本L631–648）", "PDF第7–8页选段（提取文本L743–756）", "PDF第11页选段（提取文本L1360–1377）", "PDF第5–6页选段（提取文本L482–535）", "PDF第7页选段（提取文本L718–733）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "UMI用带鱼眼相机、侧镜和IMU的手持夹爪记录示教，恢复相对末端轨迹；扩散策略配合观测同步与执行延迟匹配，减少人类采集与机器人部署的接口差异。", "experimentsZh": "四类真实任务覆盖单臂、双臂、动态及长时序操作；区分同环境能力测试与异环境泛化，并消融视野、动作表示和延迟处理。", "resultsZh": "宽视野、相对轨迹和延迟匹配共同影响成功率，不能把硬件无关接口理解为任意机器人直接可用。", "limitationsZh": "采集时不知道下游运动学限制，仍需过滤不可达动作；视觉SLAM依赖场景纹理；手持夹爪比徒手示教笨重。", "reproductionZh": "需校准相机、IMU、夹爪和机器人时钟及延迟，复现同一数据过滤与动作插值；本条未运行策略。", "experimentType": "real", "robots": [ "Universal Robots UR5", "Franka Emika FR2" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "UMI用带鱼眼相机、侧镜和IMU的手持夹爪记录示教，恢复相对末端轨迹；扩散策略配合观测同步与执行延迟匹配，减少人类采集与机器人部署的接口差异。" }, { "section": "原文实验与结果相关选段", "note": "四类真实任务覆盖单臂、双臂、动态及长时序操作；区分同环境能力测试与异环境泛化，并消融视野、动作表示和延迟处理。 宽视野、相对轨迹和延迟匹配共同影响成功率，不能把硬件无关接口理解为任意机器人直接可用。" }, { "section": "局限与复现条件", "note": "采集时不知道下游运动学限制，仍需过滤不可达动作；视觉SLAM依赖场景纹理；手持夹爪比徒手示教笨重。" } ], "originalSha256": "a9452cbaf9f5d2a77d5f8b631632fa7f45ae4bbc32400be1cc939b91a36e1609", "readingScopeSha256": "bdd7ece547f30ff5b391884db5574c8b8e58904318d760a607424f684fc1d9de", "textExtractionSha256": "84f0e4486cb33f8bd459189cbd362a54964555dada7bd9853ad445e940933b25" }, "experimentNote": "四类真实任务覆盖单臂、双臂、动态及长时序操作；区分同环境能力测试与异环境泛化，并消融视野、动作表示和延迟处理。", "contribution": "UMI用带鱼眼相机、侧镜和IMU的手持夹爪记录示教，恢复相对末端轨迹；扩散策略配合观测同步与执行延迟匹配，减少人类采集与机器人部署的接口差异。", "whyUseful": "需校准相机、IMU、夹爪和机器人时钟及延迟，复现同一数据过滤与动作插值；本条未运行策略。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "MIT；硬件、外部SLAM与数据另核。", "codeReleaseScope": "已核验官方代码与文档；未运行复现" }, { "id": "arxiv-2401.02117", "title": "Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation", "titleZh": "Mobile ALOHA：通过低成本全身遥操作学习双臂移动操作", "shortTitle": "Mobile ALOHA", "date": "2024-01-04", "year": 2024, "url": "https://arxiv.org/abs/2401.02117", "codeUrl": "https://github.com/MarkFzp/mobile-aloha", "category": "模仿学习与移动操作", "tags": [ "双臂", "移动操作", "遥操作", "ACT" ], "tier": "foundation", "summary": "将ALOHA双臂放在AgileX Tracer差速底盘上，以腰部连接反驱底盘实现手臂/移动同步遥操作。学习时把825条静态ALOHA示范与每任务移动示范等概率混合，静态数据底盘动作补零；ACT等动作块策略联合输…", "abstractZh": "将ALOHA双臂放在AgileX Tracer差速底盘上，以腰部连接反驱底盘实现手臂/移动同步遥操作。学习时把825条静态ALOHA示范与每任务移动示范等概率混合，静态数据底盘动作补零；ACT等动作块策略联合输出双臂位置与底盘速度，并补偿延迟。\n七项真实移动操作任务，每任务20–50条示范；大多数测试20回合，炒虾仅五回合。比较有无静态共训，另在擦酒/推椅任务比较ACT、Diffusion Policy和VINN，并让八位参与者学习遥操作。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "real", "robots": [ "Mobile ALOHA", "AgileX Tracer" ], "codeStatus": "open", "status": "MIT 遥操作、硬件与数据采集代码公开；训练实现另有 ACT 仓库。", "trainingNote": "每任务少量示范与静态 ALOHA 数据共训练；主代码仓库链接训练依赖，不能只装遥操作包就复现学习。", "whyUseful": "复现需移动动作归一化、静态数据补零、三视角和底盘/机械臂延迟补偿，保留任务随机化。子任务分母是进入该阶段的次数，完整任务成功为条件成功率乘积；公开硬软教程仍需独立装配与安全测试。", "contribution": "将ALOHA双臂放在AgileX Tracer差速底盘上，以腰部连接反驱底盘实现手臂/移动同步遥操作。学习时把825条静态ALOHA示范与每任务移动示范等概率混合，静态数据底盘动作补零；ACT等动作块策略联合输出双臂位置与底盘速度，并补偿延迟。", "limitations": "策略是逐任务模仿，没有自主探索或持续自我改进；示范由两位熟练操作者采集。硬件占地和固定臂高限制狭窄通道与低处操作。遥操作展示能力不能都当训练后自主策略已验证。", "caveats": "此处“全身”指双臂与轮式底盘联合控制，不是双足人形全身控制。", "evidence": [ { "url": "https://arxiv.org/abs/2401.02117", "note": "首发日期与真实任务摘要。" }, { "url": "https://arxiv.org/html/2401.02117v1", "note": "系统部分确认 AgileX Tracer 移动底盘。" }, { "url": "https://mobile-aloha.github.io/resources/mobile-aloha.pdf", "note": "图注明确自主执行使用双 ViperX 300。" }, { "url": "https://github.com/MarkFzp/mobile-aloha", "note": "MIT、遥操作代码与 ACT 训练仓库入口。" }, { "url": "https://arxiv.org/html/2401.02117v1", "note": "§3：Tracer底盘；机载3070Ti笔记本、电池和三相机。" }, { "url": "https://arxiv.org/html/2401.02117v1", "note": "§6.1 / Table 1：20次，Cook Shrimp仅5；子任务为条件成功。" }, { "url": "https://arxiv.org/html/2401.02117v1", "note": "§9：单任务模仿，两名专家示范与硬件可达性限制。" } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "运动控制", "操作与抓取", "模仿学习" ], "robotFilters": [ "Mobile ALOHA", "AgileX Tracer" ], "verificationNote": "已核论文、硬件图注与官方实现；实验为实机。", "original": { "id": "arxiv-2401.02117", "originalSourceUrl": "https://arxiv.org/pdf/2401.02117v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2401.02117v1", "finalDownloadUrl": "https://arxiv.org/pdf/2401.02117", "sha256": "ad457dae227c27b0c2a15badfea5abbc6b0802877b25feceedeb770f0ebdb604", "bytes": 4357808, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:37:52.014083+00:00", "publicationDate": "2024/01/04", "publicationDateSourceUrl": "https://arxiv.org/abs/2401.02117", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2401.02117", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2401.02117", "pages": 20, "sourceIdentityStatus": "title_match", "sourceTitle": "Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation", "sourceVersion": "2401.02117v1", "sourceVersionDate": "2024/01/04", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.012334+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2401.02117" }, "originalAnalysis": { "id": "arxiv-2401.02117", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090135+00:00", "sourceUrl": "https://arxiv.org/html/2401.02117v1", "sectionsRead": [ "§3–4", "§5–7", "Tables 1–3", "§8–9" ], "methodsZh": "将ALOHA双臂放在AgileX Tracer差速底盘上，以腰部连接反驱底盘实现手臂/移动同步遥操作。学习时把825条静态ALOHA示范与每任务移动示范等概率混合，静态数据底盘动作补零；ACT等动作块策略联合输出双臂位置与底盘速度，并补偿延迟。", "experimentsZh": "七项真实移动操作任务，每任务20–50条示范；大多数测试20回合，炒虾仅五回合。比较有无静态共训，另在擦酒/推椅任务比较ACT、Diffusion Policy和VINN，并让八位参与者学习遥操作。", "resultsZh": "作者表1擦酒完整成功由50%升至95%，冲锅0至80%，电梯0至95%；开柜仍85%，并非每项都提高。炒虾40%来自五次试验。用户练习后操作时间下降，支持界面易学但样本较小。", "limitationsZh": "策略是逐任务模仿，没有自主探索或持续自我改进；示范由两位熟练操作者采集。硬件占地和固定臂高限制狭窄通道与低处操作。遥操作展示能力不能都当训练后自主策略已验证。", "reproductionZh": "复现需移动动作归一化、静态数据补零、三视角和底盘/机械臂延迟补偿，保留任务随机化。子任务分母是进入该阶段的次数，完整任务成功为条件成功率乘积；公开硬软教程仍需独立装配与安全测试。", "experimentType": "real", "robots": [ "Mobile ALOHA", "AgileX Tracer" ], "evidenceNotes": [ { "section/page": "§3", "note": "Tracer底盘；机载3070Ti笔记本、电池和三相机。" }, { "section/page": "§6.1 / Table 1", "note": "20次，Cook Shrimp仅5；子任务为条件成功。" }, { "section/page": "§9", "note": "单任务模仿，两名专家示范与硬件可达性限制。" } ] }, "experimentNote": "七项真实移动操作任务，每任务20–50条示范；大多数测试20回合，炒虾仅五回合。比较有无静态共训，另在擦酒/推椅任务比较ACT、Diffusion Policy和VINN，并让八位参与者学习遥操作。", "analysisVerifiedAt": "2026-10-04T13:50:35.090135+00:00" }, { "id": "arxiv-2312.03275", "title": "VLFM: Vision-Language Frontier Maps for Zero-Shot Semantic Navigation", "date": "2023-12-06", "url": "https://arxiv.org/abs/2312.03275v1", "titleZh": "VLFM：面向零样本语义导航的视觉语言前沿地图", "abstractZh": "深度和里程计构建占据地图，视觉语言相似度形成目标相关价值地图；给探索前沿排序，再由低层导航器前往目标。\n在Gibson、HM3D、MP3D的ObjectNav验证集评估成功率与SPL；真实Spot用Boston Dynamics导航API，感知包括BLIP-2、检测与分割。", "summary": "深度和里程计构建占据地图，视觉语言相似度形成目标相关价值地图；给探索前沿排序，再由低层导航器前往目标。", "experimentType": "both", "robots": [ "Boston Dynamics Spot（带Spot Arm）" ], "tags": [ "导航与建图", "视觉语言动作" ], "limitations": "假定目标可从默认相机高度看到，不处理打开抽屉找物；语义价值图针对当前目标，不能直接复用于多目标连续任务。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的零样本语义导航方案，关注视觉语义地图与低层导航的分工。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2312.03275v1", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2312.03275v1", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/rai-opensource/vlfm", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/rai-opensource/vlfm", "category": "导航与建图 / 视觉语言动作", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "零样本组合导航，不训练独立导航基础模型。 使用外部 Habitat 数据，需要按数据集流程获取。 依赖外部视觉模型；本次未核实所有依赖权重的商业许可。 仓库已从 bdaiinstitute 重定向至 rai-opensource；README 说明不积极维护。开源导航代码不等于 Spot 商业软件开源。 本次未运行训练、复现实验或实机控制。", "directions": [ "导航与建图", "视觉语言动作" ], "robotFilters": [ "Boston Dynamics Spot（带Spot Arm）" ], "original": { "id": "arxiv-2312.03275", "originalSourceUrl": "https://arxiv.org/pdf/2312.03275v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2312.03275v1", "finalDownloadUrl": "https://arxiv.org/pdf/2312.03275v1", "sha256": "07f1f60fae11be46c488c1fbf5581c8f6d44904ac39ebb6071050d44d731efce", "bytes": 1653895, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:01:35.687962+00:00", "publicationDate": "2023/12/06", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2312.03275v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2312.03275v1", "pages": 7, "sourceIdentityStatus": "title_match", "sourceTitle": "VLFM: Vision-Language Frontier Maps for Zero-Shot Semantic Navigation", "sourceVersion": "2312.03275v1", "sourceVersionDate": "2023/12/06", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2312.03275v1", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2312.03275" }, "originalAnalysis": { "id": "arxiv-2312.03275", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2312.03275v1", "sourceVersion": "2312.03275v1", "sectionsRead": [ "PDF第1–2页选段（提取文本L63–80）", "PDF第5页选段（提取文本L374–387）", "PDF第5页选段（提取文本L418–435）", "PDF第6页选段（提取文本L556–569）", "PDF第6页选段（提取文本L586–603）", "PDF第2–3页选段（提取文本L164–196）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "深度和里程计构建占据地图，视觉语言相似度形成目标相关价值地图；给探索前沿排序，再由低层导航器前往目标。", "experimentsZh": "在Gibson、HM3D、MP3D的ObjectNav验证集评估成功率与SPL；真实Spot用Boston Dynamics导航API，感知包括BLIP-2、检测与分割。", "resultsZh": "比文中既有零样本方法取得更好的导航成绩；真实演示验证可用性，但低层实机导航器并非仿真PointNav策略。", "limitationsZh": "假定目标可从默认相机高度看到，不处理打开抽屉找物；语义价值图针对当前目标，不能直接复用于多目标连续任务。", "reproductionZh": "披露目标检测、单目深度与里程计来源，并区分零样本语义高层与已训练/厂商低层导航；严格保留SR和SPL定义。", "experimentType": "both", "robots": [ "Boston Dynamics Spot（带Spot Arm）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "深度和里程计构建占据地图，视觉语言相似度形成目标相关价值地图；给探索前沿排序，再由低层导航器前往目标。" }, { "section": "原文实验与结果相关选段", "note": "在Gibson、HM3D、MP3D的ObjectNav验证集评估成功率与SPL；真实Spot用Boston Dynamics导航API，感知包括BLIP-2、检测与分割。 比文中既有零样本方法取得更好的导航成绩；真实演示验证可用性，但低层实机导航器并非仿真PointNav策略。" }, { "section": "局限与复现条件", "note": "假定目标可从默认相机高度看到，不处理打开抽屉找物；语义价值图针对当前目标，不能直接复用于多目标连续任务。" } ], "originalSha256": "07f1f60fae11be46c488c1fbf5581c8f6d44904ac39ebb6071050d44d731efce", "readingScopeSha256": "b2f3e98f3eef3ed4d4e317b3ae252ba6a2130119d55e1faa85c1b5b2f0a197e8", "textExtractionSha256": "c9ec257b021de10f7bc0178ff1470961a069abae1abf8b65138bd91e330ad332" }, "experimentNote": "在Gibson、HM3D、MP3D的ObjectNav验证集评估成功率与SPL；真实Spot用Boston Dynamics导航API，感知包括BLIP-2、检测与分割。", "contribution": "深度和里程计构建占据地图，视觉语言相似度形成目标相关价值地图；给探索前沿排序，再由低层导航器前往目标。", "whyUseful": "披露目标检测、单目深度与里程计来源，并区分零样本语义高层与已训练/厂商低层导航；严格保留SR和SPL定义。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "核心 MIT；外部模型、HM3D/MP3D 数据、Spot SDK 和硬件服务独立许可。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2312.02126", "title": "SplaTAM: Splat, Track & Map 3D Gaussians for Dense RGB-D SLAM", "date": "2023-12-04", "url": "https://arxiv.org/abs/2312.02126v3", "titleZh": "SplaTAM：用三维高斯实现稠密RGB-D跟踪与建图", "abstractZh": "把场景表示为各向同性三维高斯，用可微光栅化生成颜色、深度和轮廓；交替优化相机位姿、扩展未建图区域及更新高斯参数。\n在ScanNet++、ScanNet、Replica和TUM RGB-D等序列比较位姿误差及渲染质量，区分输入视角与新视角重建。", "summary": "把场景表示为各向同性三维高斯，用可微光栅化生成颜色、深度和轮廓；交替优化相机位姿、扩展未建图区域及更新高斯参数。", "experimentType": "data", "robots": [], "tags": [ "导航与建图" ], "limitations": "对运动模糊、大深度噪声和剧烈旋转敏感；高渲染质量不自动代表全局几何一致或导航安全。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的三维高斯RGB-D SLAM研究，适合辨析跟踪、地图质量和新视角评估。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2312.02126v3", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2312.02126v3", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/spla-tam/SplaTAM", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/spla-tam/SplaTAM", "category": "导航与建图", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "发布在线地图优化、运行和评估代码。 提供数据格式、下载链接及 iPhone 采集流程。 无统一策略权重要求；地图参数按输入场景优化。 地图构建属于逐场景优化，不能称为现成通用导航策略；手机采集工具和传感器同步也需配置。 本次未运行训练、复现实验或实机控制。", "directions": [ "导航与建图" ], "robotFilters": [], "original": { "id": "arxiv-2312.02126", "originalSourceUrl": "https://arxiv.org/pdf/2312.02126v3", "versionedOriginalUrl": "https://arxiv.org/pdf/2312.02126v3", "finalDownloadUrl": "https://arxiv.org/pdf/2312.02126v3", "sha256": "c8067647cb2ba3b446017c730dfd39a453570f34e41c5b2b27d2d8a1861830fb", "bytes": 11801214, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:43.183574+00:00", "publicationDate": "2023/12/04", "license": "http://creativecommons.org/licenses/by-sa/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2312.02126v3", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2312.02126v3", "pages": 11, "sourceIdentityStatus": "title_match", "sourceTitle": "SplaTAM: Splat, Track & Map 3D Gaussians for Dense RGB-D SLAM", "sourceVersion": "2312.02126v3", "sourceVersionDate": "2024/04/16", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2312.02126v3", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2312.02126" }, "originalAnalysis": { "id": "arxiv-2312.02126", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2312.02126v3", "sourceVersion": "2312.02126v3", "sectionsRead": [ "PDF第4页选段（提取文本L298–315）", "PDF第6页选段（提取文本L549–562）", "PDF第6页选段（提取文本L586–603）", "PDF第8–9页选段（提取文本L992–1009）", "PDF第3页选段（提取文本L222–247）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "把场景表示为各向同性三维高斯，用可微光栅化生成颜色、深度和轮廓；交替优化相机位姿、扩展未建图区域及更新高斯参数。", "experimentsZh": "在ScanNet++、ScanNet、Replica和TUM RGB-D等序列比较位姿误差及渲染质量，区分输入视角与新视角重建。", "resultsZh": "论文在多个数据集改善稠密SLAM跟踪；同时明确指出只评估训练视角的渲染分数可能掩盖过拟合，应补充留出视角。", "limitationsZh": "对运动模糊、大深度噪声和剧烈旋转敏感；高渲染质量不自动代表全局几何一致或导航安全。", "reproductionZh": "固定相机内参、深度单位、帧选择及高斯增长阈值；报告ATE、几何和留出新视角指标，别只看输入帧PSNR。", "experimentType": "data", "robots": [], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "把场景表示为各向同性三维高斯，用可微光栅化生成颜色、深度和轮廓；交替优化相机位姿、扩展未建图区域及更新高斯参数。" }, { "section": "原文实验与结果相关选段", "note": "在ScanNet++、ScanNet、Replica和TUM RGB-D等序列比较位姿误差及渲染质量，区分输入视角与新视角重建。 论文在多个数据集改善稠密SLAM跟踪；同时明确指出只评估训练视角的渲染分数可能掩盖过拟合，应补充留出视角。" }, { "section": "局限与复现条件", "note": "对运动模糊、大深度噪声和剧烈旋转敏感；高渲染质量不自动代表全局几何一致或导航安全。" } ], "originalSha256": "c8067647cb2ba3b446017c730dfd39a453570f34e41c5b2b27d2d8a1861830fb", "readingScopeSha256": "0379cf68a51f77a4e530946cd320d3b0640e84ec36c9e4f16eb65747e727ac06", "textExtractionSha256": "d3ad3b7772929e48d6b44b3a5412398431900563b6f09bb613dfac8b41ca02ca" }, "experimentNote": "在ScanNet++、ScanNet、Replica和TUM RGB-D等序列比较位姿误差及渲染质量，区分输入视角与新视角重建。", "contribution": "把场景表示为各向同性三维高斯，用可微光栅化生成颜色、深度和轮廓；交替优化相机位姿、扩展未建图区域及更新高斯参数。", "whyUseful": "固定相机内参、深度单位、帧选择及高斯增长阈值；报告ATE、几何和留出新视角指标，别只看输入帧PSNR。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "主仓库 BSD-3-Clause；高斯光栅化器等依赖与各数据集单独核对，不能将整条依赖链概括为 BSD。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2311.01455", "title": "RoboGen: Towards Unleashing Infinite Data for Automated Robot Learning via Generative Simulation", "date": "2023-11-02", "url": "https://arxiv.org/abs/2311.01455v3", "titleZh": "RoboGen：通过生成式仿真自动扩展机器人技能数据", "abstractZh": "用大模型提出任务、检索或生成场景资产并校验尺寸，再生成奖励、分解步骤并选择运动规划或强化学习，形成自动技能学习流程。\n比较场景语义和尺寸校验、任务多样性与技能学习成功；刚体、关节物体、软体及运动任务都在仿真中测试。", "summary": "用大模型提出任务、检索或生成场景资产并校验尺寸，再生成奖励、分解步骤并选择运动规划或强化学习，形成自动技能学习流程。", "experimentType": "sim", "robots": [ "Franka Panda（仿真示例）", "腿式仿真平台（完整型号未核实）" ], "tags": [ "数据集与基准", "强化学习", "世界模型" ], "limitations": "大规模技能验证仍困难，物体几何/关节语义和精密配合易出错；真实部署还受仿真到现实差距限制。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的生成任务和仿真技能学习系统，关注自动验证及公开实现覆盖范围。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2311.01455v3", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2311.01455v3", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/Genesis-Embodied-AI/RoboGen", "note": "公开的是PyBullet刚体操作与运动任务重实现；非论文原始Genesis及软体完整流程；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/Genesis-Embodied-AI/RoboGen", "category": "数据集与基准 / 强化学习", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "已核验任务生成与技能训练入口，含SAC和CEM。README明确本仓库为PyBullet重实现，覆盖刚体操作和运动任务；原论文内部Genesis版本及软体流程并未由此完整公开。预生成任务、资产和嵌入需遵守原来源许可；未核实统一预训练通用策略包。 源码开放不能解释为论文全部实验可完整复现；自动任务、几何与奖励仍需独立验证。 本次未运行训练、复现实验或实机控制。", "directions": [ "数据集与基准", "强化学习", "世界模型" ], "robotFilters": [ "Franka Panda（仿真示例）", "腿式仿真平台（完整型号未核实）" ], "original": { "id": "arxiv-2311.01455", "originalSourceUrl": "https://arxiv.org/pdf/2311.01455v3", "versionedOriginalUrl": "https://arxiv.org/pdf/2311.01455v3", "finalDownloadUrl": "https://arxiv.org/pdf/2311.01455v3", "sha256": "4146b2443205d6e4f8137e14fb20a91165da9a70c12306a300bc71fcae53b022", "bytes": 34123085, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:01:12.102944+00:00", "publicationDate": "2023/11/02", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2311.01455v3", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2311.01455v3", "pages": 48, "sourceIdentityStatus": "title_match", "sourceTitle": "RoboGen: Towards Unleashing Infinite Data for Automated Robot Learning via Generative Simulation", "sourceVersion": "2311.01455v3", "sourceVersionDate": "2024/06/14", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2311.01455v3", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2311.01455" }, "originalAnalysis": { "id": "arxiv-2311.01455", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2311.01455v3", "sourceVersion": "2311.01455v3", "sectionsRead": [ "PDF第3页选段（提取文本L212–225）", "PDF第6页选段（提取文本L439–456）", "PDF第7页选段（提取文本L576–589）", "PDF第6–7页选段（提取文本L465–500）", "PDF第8页选段（提取文本L594–621）", "PDF第3页选段（提取文本L175–209）", "PDF第9页选段（提取文本L708–728）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "用大模型提出任务、检索或生成场景资产并校验尺寸，再生成奖励、分解步骤并选择运动规划或强化学习，形成自动技能学习流程。", "experimentsZh": "比较场景语义和尺寸校验、任务多样性与技能学习成功；刚体、关节物体、软体及运动任务都在仿真中测试。", "resultsZh": "在关节物体任务上结合规划原语比纯RL更容易学成；展示100多种技能，155项任务人工检查仍发现失败，不能称作无限可靠数据。", "limitationsZh": "大规模技能验证仍困难，物体几何/关节语义和精密配合易出错；真实部署还受仿真到现实差距限制。", "reproductionZh": "原论文用了内部Genesis版本；已公开PyBullet实现仅覆盖部分流程，应单独核实软体、资产和模型依赖，而非宣称全部可复现。", "experimentType": "sim", "robots": [ "Franka Panda（仿真示例）", "腿式仿真平台（完整型号未核实）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "用大模型提出任务、检索或生成场景资产并校验尺寸，再生成奖励、分解步骤并选择运动规划或强化学习，形成自动技能学习流程。" }, { "section": "原文实验与结果相关选段", "note": "比较场景语义和尺寸校验、任务多样性与技能学习成功；刚体、关节物体、软体及运动任务都在仿真中测试。 在关节物体任务上结合规划原语比纯RL更容易学成；展示100多种技能，155项任务人工检查仍发现失败，不能称作无限可靠数据。" }, { "section": "局限与复现条件", "note": "大规模技能验证仍困难，物体几何/关节语义和精密配合易出错；真实部署还受仿真到现实差距限制。" } ], "originalSha256": "4146b2443205d6e4f8137e14fb20a91165da9a70c12306a300bc71fcae53b022", "readingScopeSha256": "ee5fc411547e98734cecd60687acdd00f80d8e06ce4357654d1e64b81e98fc60", "textExtractionSha256": "766a1dca6019db4a59559b93bf683e2fc2225b0518994953f4c5c5ddd09d4f62" }, "experimentNote": "比较场景语义和尺寸校验、任务多样性与技能学习成功；刚体、关节物体、软体及运动任务都在仿真中测试。", "contribution": "用大模型提出任务、检索或生成场景资产并校验尺寸，再生成奖励、分解步骤并选择运动规划或强化学习，形成自动技能学习流程。", "whyUseful": "原论文用了内部Genesis版本；已公开PyBullet实现仅覆盖部分流程，应单独核实软体、资产和模型依赖，而非宣称全部可复现。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "代码 Apache-2.0；PartNet-Mobility、Objaverse 资产和外部模型/API 条款单独适用。", "codeReleaseScope": "公开的是PyBullet刚体操作与运动任务重实现；非论文原始Genesis及软体完整流程" }, { "id": "arxiv-2310.17596", "title": "MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations", "date": "2023-10-26", "url": "https://arxiv.org/abs/2310.17596v1", "titleZh": "MimicGen：由少量人类示教生成大规模机器人学习数据", "abstractZh": "把示教切成以物体为中心的子任务，再按新场景的物体位姿变换末端轨迹，插值连接片段并过滤未成功的生成轨迹。\n比较初态分布、同类别物体和机器人变体；论文包含仿真及真实机械臂实验，并用生成数据训练图像输入的BC-RNN。", "summary": "把示教切成以物体为中心的子任务，再按新场景的物体位姿变换末端轨迹，插值连接片段并过滤未成功的生成轨迹。", "experimentType": "both", "robots": [ "Franka Panda（仿真）", "Rethink Sawyer（仿真）", "KUKA IIWA（仿真）", "Universal Robots UR5e（仿真）", "实机型号未核实" ], "tags": [ "数据集与基准", "模仿学习", "操作与抓取" ], "limitations": "需要子任务划分、物体位姿及成功判据；线性插值不保证无碰撞；成功过滤会造成数据偏差，主要验证刚性物体的准静态任务。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的示教扩增方法，用于理解物体相对轨迹重用和数据生成偏差。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "源码可用，存在使用限制", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2310.17596v1", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2310.17596v1", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/NVlabs/mimicgen", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "source_available", "codeUrl": "https://github.com/NVlabs/mimicgen", "category": "数据集与基准 / 模仿学习", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "已发布生成代码、配置生成器与 robomimic 训练命令。 官方文档提供 12 任务、超过 48,000 条示范及 Hugging Face 下载。 本次未核实独立预训练策略权重下载。 公开环境从 robosuite 1.2 迁至 1.4，外观和动力学差异可能影响复现；实机控制栈不等同于这套仿真发布。 本次未运行训练、复现实验或实机控制。", "directions": [ "数据集与基准", "模仿学习", "操作与抓取" ], "robotFilters": [ "Franka Panda（仿真）", "Rethink Sawyer（仿真）", "KUKA IIWA（仿真）", "Universal Robots UR5e（仿真）", "实机型号未核实" ], "original": { "id": "arxiv-2310.17596", "originalSourceUrl": "https://arxiv.org/pdf/2310.17596v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2310.17596v1", "finalDownloadUrl": "https://arxiv.org/pdf/2310.17596v1", "sha256": "7345661fb8e8581981e272e1e043561cbfdaf3f8403d2cb5e28f70797c9bbb16", "bytes": 14209120, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T14:59:32.240117+00:00", "publicationDate": "2023/10/26", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2310.17596v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2310.17596v1", "pages": 45, "sourceIdentityStatus": "title_match", "sourceTitle": "MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations", "sourceVersion": "2310.17596v1", "sourceVersionDate": "2023/10/26", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2310.17596v1", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2310.17596" }, "originalAnalysis": { "id": "arxiv-2310.17596", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2310.17596v1", "sourceVersion": "2310.17596v1", "sectionsRead": [ "PDF第4页选段（提取文本L273–290）", "PDF第5页选段（提取文本L331–344）", "PDF第6页选段（提取文本L535–548）", "PDF第8页选段（提取文本L658–675）", "PDF第3–4页选段（提取文本L180–213）", "PDF第7页选段（提取文本L572–610）", "PDF第8页选段（提取文本L637–658）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "把示教切成以物体为中心的子任务，再按新场景的物体位姿变换末端轨迹，插值连接片段并过滤未成功的生成轨迹。", "experimentsZh": "比较初态分布、同类别物体和机器人变体；论文包含仿真及真实机械臂实验，并用生成数据训练图像输入的BC-RNN。", "resultsZh": "不足200条人类示教生成逾5万条轨迹、覆盖18项任务；实机生成成功不等于策略成功：Stack和Coffee策略各50次测试分别成功36%与14%。", "limitationsZh": "需要子任务划分、物体位姿及成功判据；线性插值不保证无碰撞；成功过滤会造成数据偏差，主要验证刚性物体的准静态任务。", "reproductionZh": "应分别报告生成成功率与最终策略成功率，复现来源片段选择和评估选点；论文报告多个检查点中的最佳成功率。", "experimentType": "both", "robots": [ "Franka Panda（仿真）", "Rethink Sawyer（仿真）", "KUKA IIWA（仿真）", "Universal Robots UR5e（仿真）", "实机型号未核实" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "把示教切成以物体为中心的子任务，再按新场景的物体位姿变换末端轨迹，插值连接片段并过滤未成功的生成轨迹。" }, { "section": "原文实验与结果相关选段", "note": "比较初态分布、同类别物体和机器人变体；论文包含仿真及真实机械臂实验，并用生成数据训练图像输入的BC-RNN。 不足200条人类示教生成逾5万条轨迹、覆盖18项任务；实机生成成功不等于策略成功：Stack和Coffee策略各50次测试分别成功36%与14%。" }, { "section": "局限与复现条件", "note": "需要子任务划分、物体位姿及成功判据；线性插值不保证无碰撞；成功过滤会造成数据偏差，主要验证刚性物体的准静态任务。" } ], "originalSha256": "7345661fb8e8581981e272e1e043561cbfdaf3f8403d2cb5e28f70797c9bbb16", "readingScopeSha256": "6fc1956bdda05667ddaf92c6156d4468e17f541b9109e0dd591c9eb2d41b05cf", "textExtractionSha256": "b3d58d168058f876da52172c74d680eac27c62ca33c88bd03428290342cd07a4" }, "experimentNote": "比较初态分布、同类别物体和机器人变体；论文包含仿真及真实机械臂实验，并用生成数据训练图像输入的BC-RNN。", "contribution": "把示教切成以物体为中心的子任务，再按新场景的物体位姿变换末端轨迹，插值连接片段并过滤未成功的生成轨迹。", "whyUseful": "应分别报告生成成功率与最终策略成功率，复现来源片段选择和评估选点；论文报告多个检查点中的最佳成功率。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "代码为 NVIDIA Source Code License，含非商业限制；发布数据为 CC BY 4.0。两者不能混用许可结论。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2310.16828", "title": "TD-MPC2: Scalable, Robust World Models for Continuous Control", "titleZh": "TD-MPC2：面向连续控制的可扩展、稳健世界模型", "shortTitle": "TD-MPC2", "date": "2023-10-25", "year": 2023, "url": "https://arxiv.org/abs/2310.16828", "paperUrl": "https://arxiv.org/abs/2310.16828", "projectUrl": "https://www.tdmpc2.com/", "codeUrl": "https://github.com/nicklashansen/tdmpc2", "summary": "学习不解码图像的控制导向潜在动力学，同时预测奖励和TD终值，再用短时域采样MPC结合终值与策略先验选动作。SimNorm把潜在维度分组归一化，奖励/价值使用对数空间离散回归，多任务借任务嵌入、状态填充和动作掩码…", "abstractZh": "学习不解码图像的控制导向潜在动力学，同时预测奖励和TD终值，再用短时域采样MPC结合终值与策略先验选动作。SimNorm把潜在维度分组归一化，奖励/价值使用对数空间离散回归，多任务借任务嵌入、状态填充和动作掩码共享模型。\n在DMControl39、Meta-World50、ManiSkill2五及MyoSuite十任务共104个仿真连续控制任务上，以同组超参数对比SAC、DreamerV3和TD-MPC，主要结果三种子及95%区间；另测试80任务离线数据预训练、模型规模和新任务微调。", "category": "世界模型 / 多任务控制", "tags": [ "TD-MPC2", "世界模型", "强化学习" ], "directions": [ "世界模型", "强化学习" ], "tier": "classic", "experimentType": "sim", "experimentNote": "在DMControl39、Meta-World50、ManiSkill2五及MyoSuite十任务共104个仿真连续控制任务上，以同组超参数对比SAC、DreamerV3和TD-MPC，主要结果三种子及95%区间；另测试80任务离线数据预训练、模型规模和新任务微调。", "robots": [], "robotFilters": [], "robotNote": "无实机；仿真形态不等同于已验证的商业机器人型号。", "codeStatus": "open", "status": "训练代码已开源", "trainingStatus": "训练代码已开源", "trainingNote": "官方 MIT 仓库提供训练、模型和数据链接；README 明确第三方代码遵守各自许可。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "MIT（主代码；第三方另行许可）", "contribution": "学习不解码图像的控制导向潜在动力学，同时预测奖励和TD终值，再用短时域采样MPC结合终值与策略先验选动作。SimNorm把潜在维度分组归一化，奖励/价值使用对数空间离散回归，多任务借任务嵌入、状态填充和动作掩码共享模型。", "whyUseful": "应同时复现奖励尺度处理、Q集成、SimNorm、动作掩码和规划预算，不能只替换骨干后归因容量。需区分单任务在线交互和离线多任务训练的数据预算，本次未运行基准。", "limitations": "主要使用状态输入和仿真，像素结果只在子集验证，没有真实机器人部署证据。MPC仍依赖模型和价值误差，模型变大增加推理/训练成本；跨任务共享需要任务标识，非自然语言开放目标理解。", "caveats": "主要证据来自仿真；仍依赖奖励、交互和在线规划，不能直接保证实机安全或域外泛化。", "evidence": [ { "url": "https://arxiv.org/abs/2310.16828", "note": "首发日期与规模实验。" }, { "url": "https://www.tdmpc2.com/", "note": "104 任务、四类仿真域及多任务模型。" }, { "url": "https://github.com/nicklashansen/tdmpc2", "note": "官方实现、训练和 MIT/第三方许可说明。" }, { "url": "https://arxiv.org/pdf/2310.16828", "note": "§3; Figure 4：控制导向潜在目标、104仿真任务三种子。" }, { "url": "https://arxiv.org/pdf/2310.16828", "note": "§4 Figure 7; §5：1M至317M规模、多任务归一化分数与部署限制。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2023：从单任务模型控制推进到大规模多任务、跨本体潜空间规划。", "freshness": "经典工作", "original": { "id": "arxiv-2310.16828", "originalSourceUrl": "https://arxiv.org/pdf/2310.16828v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2310.16828v2", "finalDownloadUrl": "https://arxiv.org/pdf/2310.16828", "sha256": "99832c555b66665ad1e26cd48e75ea9dda4f3750175b993e7ed5b8dfe399b9e0", "bytes": 7858004, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:37:56.339041+00:00", "publicationDate": "2023/10/25", "publicationDateSourceUrl": "https://arxiv.org/abs/2310.16828", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2310.16828", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2310.16828", "pages": 31, "sourceIdentityStatus": "title_match", "sourceTitle": "TD-MPC2: Scalable, Robust World Models for Continuous Control", "sourceVersion": "2310.16828v2", "sourceVersionDate": "2024/03/21", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.018684+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2310.16828" }, "originalAnalysis": { "id": "arxiv-2310.16828", "sourceUrl": "https://arxiv.org/pdf/2310.16828", "sectionsRead": [ "§3.1–3.3", "§4", "Figures 4,7", "§5" ], "methodsZh": "学习不解码图像的控制导向潜在动力学，同时预测奖励和TD终值，再用短时域采样MPC结合终值与策略先验选动作。SimNorm把潜在维度分组归一化，奖励/价值使用对数空间离散回归，多任务借任务嵌入、状态填充和动作掩码共享模型。", "experimentsZh": "在DMControl39、Meta-World50、ManiSkill2五及MyoSuite十任务共104个仿真连续控制任务上，以同组超参数对比SAC、DreamerV3和TD-MPC，主要结果三种子及95%区间；另测试80任务离线数据预训练、模型规模和新任务微调。", "resultsZh": "作者报告总体数据效率与最终表现优于对照，规模实验在80任务上由1M模型约16的归一化分数升至317M模型约70.6。归一化分数并非统一成功率；结果显示模型容量、SimNorm及价值学习稳定化共同影响表现。", "limitationsZh": "主要使用状态输入和仿真，像素结果只在子集验证，没有真实机器人部署证据。MPC仍依赖模型和价值误差，模型变大增加推理/训练成本；跨任务共享需要任务标识，非自然语言开放目标理解。", "reproductionZh": "应同时复现奖励尺度处理、Q集成、SimNorm、动作掩码和规划预算，不能只替换骨干后归因容量。需区分单任务在线交互和离线多任务训练的数据预算，本次未运行基准。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "§3; Figure 4", "note": "控制导向潜在目标、104仿真任务三种子。" }, { "section/page": "§4 Figure 7; §5", "note": "1M至317M规模、多任务归一化分数与部署限制。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2310.13724", "title": "Habitat 3.0: A Co-Habitat for Humans, Avatars and Robots", "date": "2023-10-19", "url": "https://arxiv.org/abs/2310.13724v1", "titleZh": "Habitat 3.0：人类、虚拟人和机器人的协作居家环境", "abstractZh": "将骨架碰撞、蒙皮外观和SMPL-X人体模型集成到仿真，配合行走/伸手行为库和人在环控制工具研究跟随与协作重排。\nSpot仿真在HSSD场景测试社会导航与零样本伙伴协作；另有30位参与者的人在环仿真实验，不是实机人机共处试验。", "summary": "将骨架碰撞、蒙皮外观和SMPL-X人体模型集成到仿真，配合行走/伸手行为库和人在环控制工具研究跟随与协作重排。", "experimentType": "sim", "robots": [ "Boston Dynamics Spot（仿真）" ], "tags": [ "数据集与基准", "导航与建图", "多机器人" ], "limitations": "人体主要只做行走和伸手；部分重排输入含目标/初始物体真值，开放容器假设较强，人在环观测与虚拟人不同。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "foundation", "tierNote": "编辑分类：人机协作仿真与人在环评估基础设施；分类依据工具角色，不依据发表年份。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2310.13724v1", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2310.13724v1", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/facebookresearch/habitat-lab", "note": "已核验官方代码与文档；未运行复现；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/facebookresearch/habitat-lab", "category": "数据集与基准 / 导航与建图", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "README已核实人类跟随、重排、人在环工具与多智能体训练；原论文伙伴群体和理想技能设置需按原配置复现。 本次未运行训练、复现实验或实机控制。", "directions": [ "数据集与基准", "导航与建图", "多机器人" ], "robotFilters": [ "Boston Dynamics Spot（仿真）" ], "original": { "id": "arxiv-2310.13724", "originalSourceUrl": "https://arxiv.org/pdf/2310.13724v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2310.13724v1", "finalDownloadUrl": "https://arxiv.org/pdf/2310.13724v1", "sha256": "6383018901acbf7677050585da9a30707437cf7119def797be500f33a68021af", "bytes": 8852187, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:09.351040+00:00", "publicationDate": "2023/10/19", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2310.13724v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2310.13724v1", "pages": 30, "sourceIdentityStatus": "title_match", "sourceTitle": "Habitat 3.0: A Co-Habitat for Humans, Avatars and Robots", "sourceVersion": "2310.13724v1", "sourceVersionDate": "2023/10/19", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2310.13724v1", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2310.13724" }, "originalAnalysis": { "id": "arxiv-2310.13724", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2310.13724v1", "sourceVersion": "2310.13724v1", "sectionsRead": [ "PDF第2页选段（提取文本L71–88）", "PDF第4页选段（提取文本L207–224）", "PDF第7页选段（提取文本L423–436）", "PDF第7–8页选段（提取文本L462–479）", "PDF第9页选段（提取文本L630–665）", "PDF第30页选段（提取文本L2605–2635）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "将骨架碰撞、蒙皮外观和SMPL-X人体模型集成到仿真，配合行走/伸手行为库和人在环控制工具研究跟随与协作重排。", "experimentsZh": "Spot仿真在HSSD场景测试社会导航与零样本伙伴协作；另有30位参与者的人在环仿真实验，不是实机人机共处试验。", "resultsZh": "伙伴群体训练有利于对新伙伴协作；将理想底层技能换成学习技能后任务成功明显下降，说明高层协作依赖执行可靠性。", "limitationsZh": "人体主要只做行走和伸手；部分重排输入含目标/初始物体真值，开放容器假设较强，人在环观测与虚拟人不同。", "reproductionZh": "分别记录理想技能和学习技能、伙伴训练/测试群体及人体传感器；不要把仿真人参与研究标记为真实机器人验证。", "experimentType": "sim", "robots": [ "Boston Dynamics Spot（仿真）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "将骨架碰撞、蒙皮外观和SMPL-X人体模型集成到仿真，配合行走/伸手行为库和人在环控制工具研究跟随与协作重排。" }, { "section": "原文实验与结果相关选段", "note": "Spot仿真在HSSD场景测试社会导航与零样本伙伴协作；另有30位参与者的人在环仿真实验，不是实机人机共处试验。 伙伴群体训练有利于对新伙伴协作；将理想底层技能换成学习技能后任务成功明显下降，说明高层协作依赖执行可靠性。" }, { "section": "局限与复现条件", "note": "人体主要只做行走和伸手；部分重排输入含目标/初始物体真值，开放容器假设较强，人在环观测与虚拟人不同。" } ], "originalSha256": "6383018901acbf7677050585da9a30707437cf7119def797be500f33a68021af", "readingScopeSha256": "abcb3d675affc00fb754e6a5084c0d4f075c3f190ba63243515d926421e7ffb4", "textExtractionSha256": "ff8e788eac37dd77419f4e503175b016e4d702b230718755b8630dcbcf1ac3c6" }, "experimentNote": "Spot仿真在HSSD场景测试社会导航与零样本伙伴协作；另有30位参与者的人在环仿真实验，不是实机人机共处试验。", "contribution": "将骨架碰撞、蒙皮外观和SMPL-X人体模型集成到仿真，配合行走/伸手行为库和人在环控制工具研究跟随与协作重排。", "whyUseful": "分别记录理想技能和学习技能、伙伴训练/测试群体及人体传感器；不要把仿真人参与研究标记为真实机器人验证。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "核心MIT；人体模型、场景和依赖单独核对。", "codeReleaseScope": "已核验官方代码与文档；未运行复现" }, { "id": "arxiv-2310.08864", "title": "Open X-Embodiment: Robotic Learning Datasets and RT-X Models", "titleZh": "Open X-Embodiment：机器人学习数据集与 RT-X 模型", "shortTitle": "Open X / RT-X", "date": "2023-10-13", "year": 2023, "url": "https://arxiv.org/abs/2310.08864", "codeUrl": "https://github.com/google-deepmind/open_x_embodiment", "category": "数据集与基准", "tags": [ "跨本体", "RLDS", "RT-1-X", "RT-2-X" ], "tier": "foundation", "summary": "Open X-Embodiment把跨机构机器人轨迹统一为RLDS接口，以RT-1-X及RT-2-X检验跨形态联合训练。图像和语言统一处理，动作粗略归一化但不强制对齐坐标、绝对/相对或速度语义；大模型将动作当离…", "abstractZh": "Open X-Embodiment把跨机构机器人轨迹统一为RLDS接口，以RT-1-X及RT-2-X检验跨形态联合训练。图像和语言统一处理，动作粗略归一化但不强制对齐坐标、绝对/相对或速度语义；大模型将动作当离散语言词元与网页视觉语言知识结合。\n发布目录覆盖22形态、21机构；论文训练混合实际只取9种机械臂，并在六机器人做3600次真实测试。小数据域研究同任务迁移，大数据域比较模型容量，再在Google Robot上测试原本只出现在WidowX Bridge数据里的技能。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "real", "robots": [ "Google Robot", "WidowX", "UR5", "多平台（六种评测形态）" ], "codeStatus": "open", "status": "数据工具与 RT-1-X 推理代码、检查点公开。", "trainingNote": "软件 Apache-2.0；各来源数据应单独核许可证。开放 RT-1-X 不代表 RT-2-X 全部训练代码和权重开放。", "whyUseful": "先明确9形态训练混合及采样权重，逐数据集核对动作单位和控制频率；对同一架构比较单域与混合，再做移除Bridge消融。分开汇报新技能、常规泛化及小/大数据域，避免合并成通用“提升三倍”。", "contribution": "Open X-Embodiment把跨机构机器人轨迹统一为RLDS接口，以RT-1-X及RT-2-X检验跨形态联合训练。图像和语言统一处理，动作粗略归一化但不强制对齐坐标、绝对/相对或速度语义；大模型将动作当离散语言词元与网页视觉语言知识结合。", "limitations": "未研究完全新机器人零样本迁移，也不覆盖截然不同感知/动作模态；正迁移依赖容量、数据组成及网络预训练。目录有22形态不等于模型训练和评测用了全部22种。", "caveats": "22 种数据本体、9 种训练本体及评测机器人数量是不同统计口径。", "evidence": [ { "url": "https://arxiv.org/abs/2310.08864", "note": "标题、摘要与首发日期。" }, { "url": "https://arxiv.org/html/2310.08864v1", "note": "第 III–V 节明确数据、训练与实机评测范围。" }, { "url": "https://github.com/google-deepmind/open_x_embodiment", "note": "RLDS 工具、RT-1-X 检查点与许可证。" }, { "url": "https://arxiv.org/pdf/2310.08864", "note": "IV; p.5 Fig.4/Table I（已渲染）：训练形态数、真实评估与63/41/44均值。" }, { "url": "https://arxiv.org/pdf/2310.08864", "note": "Table II; VI：新技能75.8对27.3、常规泛化持平及未见形态边界。" }, { "url": "https://arxiv.org/pdf/2310.08864v9", "note": "IV; p.5 Fig.4/Table I（已渲染）：训练形态数、真实评估与63/41/44均值。" }, { "url": "https://arxiv.org/pdf/2310.08864v9", "note": "Table II; VI：新技能75.8对27.3、常规泛化持平及未见形态边界。" } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "视觉语言动作", "数据集与基准" ], "robotFilters": [ "Google Robot", "WidowX", "Universal Robots UR5" ], "verificationNote": "已核原论文及官方仓库；机器人列表列出已核代表评测平台，非全部数据本体。", "original": { "id": "arxiv-2310.08864", "originalSourceUrl": "https://arxiv.org/pdf/2310.08864v9", "versionedOriginalUrl": "https://arxiv.org/pdf/2310.08864v9", "finalDownloadUrl": "https://arxiv.org/pdf/2310.08864", "sha256": "13f16aff5afdee583dc25b3c570e9bb0cf6baefb59b539928d919588cd8d9246", "bytes": 4011306, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:37:56.305186+00:00", "publicationDate": "2023/10/13", "publicationDateSourceUrl": "https://arxiv.org/abs/2310.08864", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2310.08864", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2310.08864", "pages": 12, "sourceIdentityStatus": "title_match", "sourceTitle": "Open X-Embodiment: Robotic Learning Datasets and RT-X Models", "sourceVersion": "2310.08864v9", "sourceVersionDate": "2025/05/14", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.022749+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2310.08864" }, "originalAnalysis": { "id": "arxiv-2310.08864", "sourceUrl": "https://arxiv.org/pdf/2310.08864v9", "sectionsRead": [ "III", "IV-A–C", "V-A–C", "VI", "Fig.4（渲染核对）; Tables I–II" ], "methodsZh": "Open X-Embodiment把跨机构机器人轨迹统一为RLDS接口，以RT-1-X及RT-2-X检验跨形态联合训练。图像和语言统一处理，动作粗略归一化但不强制对齐坐标、绝对/相对或速度语义；大模型将动作当离散语言词元与网页视觉语言知识结合。", "experimentsZh": "发布目录覆盖22形态、21机构；论文训练混合实际只取9种机械臂，并在六机器人做3600次真实测试。小数据域研究同任务迁移，大数据域比较模型容量，再在Google Robot上测试原本只出现在WidowX Bridge数据里的技能。", "resultsZh": "图4 RT-1-X小数据域平均63%，原方法41%、单域RT-1为44%，约50%相对提升而非50个百分点。55B RT-2-X新技能75.8%，RT-2为27.3%；但普通对象/背景泛化61%对62%，未提升。小RT-1-X在大数据域可能欠拟合。", "limitationsZh": "未研究完全新机器人零样本迁移，也不覆盖截然不同感知/动作模态；正迁移依赖容量、数据组成及网络预训练。目录有22形态不等于模型训练和评测用了全部22种。", "reproductionZh": "先明确9形态训练混合及采样权重，逐数据集核对动作单位和控制频率；对同一架构比较单域与混合，再做移除Bridge消融。分开汇报新技能、常规泛化及小/大数据域，避免合并成通用“提升三倍”。", "experimentType": "real", "robots": [ "Google Robot", "WidowX", "UR5", "多平台（六种评测形态）" ], "corrections": [ "22形态是资源目录覆盖，实验训练混合为9种；50%为相对提升，不是百分点。" ], "evidenceNotes": [ { "note": "训练形态数、真实评估与63/41/44均值。", "section": "IV; p.5 Fig.4/Table I（已渲染）" }, { "note": "新技能75.8对27.3、常规泛化持平及未见形态边界。", "section": "Table II; VI" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2310.08864v9", "originalSha256": "13f16aff5afdee583dc25b3c570e9bb0cf6baefb59b539928d919588cd8d9246" }, "experimentNote": "发布目录覆盖22形态、21机构；论文训练混合实际只取9种机械臂，并在六机器人做3600次真实测试。小数据域研究同任务迁移，大数据域比较模型容量，再在Google Robot上测试原本只出现在WidowX Bridge数据里的技能。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2310.08576", "title": "Learning to Act from Actionless Videos through Dense Correspondences", "date": "2023-10-12", "url": "https://arxiv.org/abs/2310.08576v1", "titleZh": "AVDC：通过稠密对应从无动作标签视频学习行为", "abstractZh": "先生成任务视频，用GMFlow估计逐帧稠密光流，再结合初始深度恢复物体或相机的三维刚体运动，由现有操作原语执行并在停滞时重新规划。\n评估Meta-World操作、iTHOR导航、人类推物视频迁移及真实Panda；实机Bridge预训练后还用20段本场景人类示范视频微调。", "summary": "先生成任务视频，用GMFlow估计逐帧稠密光流，再结合初始深度恢复物体或相机的三维刚体运动，由现有操作原语执行并在停滞时重新规划。", "experimentType": "both", "robots": [ "Franka Emika Panda", "Trossen WidowX 250（源数据平台）" ], "tags": [ "世界模型", "操作与抓取", "导航与建图" ], "limitations": "遮挡、光照急变和大运动会破坏光流；实机假设可顶抓且无需重定向物体，RGB视频也无法提供接触力信息。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的无动作视频到控制研究，关注光流、深度及实机适配条件。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2310.08576v1", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2310.08576v1", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/flow-diffusion/AVDC", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/flow-diffusion/AVDC", "category": "世界模型 / 操作与抓取", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "提供 Meta-World、iTHOR、Bridge 的训练和推理脚本。 提供数据目录规范；本次未核实训练全集可公开下载。 作者提供三个实验域的 Hugging Face 检查点。 主仓库的视频预测权重不等于完整动作执行链；迁移真实相机和机械臂需重新校准对应与控制。 本次未运行训练、复现实验或实机控制。", "directions": [ "世界模型", "操作与抓取", "导航与建图" ], "robotFilters": [ "Franka Emika Panda", "Trossen WidowX 250（源数据平台）" ], "original": { "id": "arxiv-2310.08576", "originalSourceUrl": "https://arxiv.org/pdf/2310.08576v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2310.08576v1", "finalDownloadUrl": "https://arxiv.org/pdf/2310.08576v1", "sha256": "678d6b56c8acf054a0af49c287534c43ad7b0a7ffb3f00153512a1d41b3df540", "bytes": 11181162, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:01:05.977138+00:00", "publicationDate": "2023/10/12", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2310.08576v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2310.08576v1", "pages": 19, "sourceIdentityStatus": "title_match", "sourceTitle": "Learning to Act from Actionless Videos through Dense Correspondences", "sourceVersion": "2310.08576v1", "sourceVersionDate": "2023/10/12", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2310.08576v1", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2310.08576" }, "originalAnalysis": { "id": "arxiv-2310.08576", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2310.08576v1", "sourceVersion": "2310.08576v1", "sectionsRead": [ "PDF第5页选段（提取文本L377–394）", "PDF第6–7页选段（提取文本L450–463）", "PDF第8页选段（提取文本L783–800）", "PDF第9页选段（提取文本L850–867）", "PDF第3–4页选段（提取文本L232–278）", "PDF第8–9页选段（提取文本L804–841）", "PDF第4页选段（提取文本L279–307）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "先生成任务视频，用GMFlow估计逐帧稠密光流，再结合初始深度恢复物体或相机的三维刚体运动，由现有操作原语执行并在停滞时重新规划。", "experimentsZh": "评估Meta-World操作、iTHOR导航、人类推物视频迁移及真实Panda；实机Bridge预训练后还用20段本场景人类示范视频微调。", "resultsZh": "完整方法优于文中无重规划与直接光流生成变体；真实Panda结果不是仅凭Bridge零样本部署，必须保留20段微调数据的条件。", "limitationsZh": "遮挡、光照急变和大运动会破坏光流；实机假设可顶抓且无需重定向物体，RGB视频也无法提供接触力信息。", "reproductionZh": "核对分割、深度、光流与坐标变换，分别标记源机器人数据和实机微调；保留抓取原语与重规划触发条件。", "experimentType": "both", "robots": [ "Franka Emika Panda", "Trossen WidowX 250（源数据平台）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "先生成任务视频，用GMFlow估计逐帧稠密光流，再结合初始深度恢复物体或相机的三维刚体运动，由现有操作原语执行并在停滞时重新规划。" }, { "section": "原文实验与结果相关选段", "note": "评估Meta-World操作、iTHOR导航、人类推物视频迁移及真实Panda；实机Bridge预训练后还用20段本场景人类示范视频微调。 完整方法优于文中无重规划与直接光流生成变体；真实Panda结果不是仅凭Bridge零样本部署，必须保留20段微调数据的条件。" }, { "section": "局限与复现条件", "note": "遮挡、光照急变和大运动会破坏光流；实机假设可顶抓且无需重定向物体，RGB视频也无法提供接触力信息。" } ], "originalSha256": "678d6b56c8acf054a0af49c287534c43ad7b0a7ffb3f00153512a1d41b3df540", "readingScopeSha256": "e68e16a6e1f4b329d7b5654c7449097026f18fcb10609f7f8a82e1f1d52b010d", "textExtractionSha256": "3f2bff68c9f8c317921d56dd7d678fa368cc7f37acc91928aa1b37a15e47ed53" }, "experimentNote": "评估Meta-World操作、iTHOR导航、人类推物视频迁移及真实Panda；实机Bridge预训练后还用20段本场景人类示范视频微调。", "contribution": "先生成任务视频，用GMFlow估计逐帧稠密光流，再结合初始深度恢复物体或相机的三维刚体运动，由现有操作原语执行并在停滞时重新规划。", "whyUseful": "核对分割、深度、光流与坐标变换，分别标记源机器人数据和实机微调；保留抓取原语与重规划触发条件。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "主仓库 MIT；视频数据、模型权重与实验依赖按各自许可。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2310.07896", "title": "NoMaD: Goal Masked Diffusion Policies for Navigation and Exploration", "date": "2023-10-11", "url": "https://arxiv.org/abs/2310.07896v1", "titleZh": "NoMaD：统一探索与导航的目标掩码扩散策略", "abstractZh": "在ViNT式视觉编码中随机遮蔽目标token，用同一个动作扩散模型学习有目标导航与无目标探索，配合拓扑图进行长时序搜索。\n六种真实室内外环境，比较未知环境探索和已知环境图像目标导航；同时报告成功率与每次试验的碰撞数。", "summary": "在ViNT式视觉编码中随机遮蔽目标token，用同一个动作扩散模型学习有目标导航与无目标探索，配合拓扑图进行长时序搜索。", "experimentType": "real", "robots": [ "LoCoBot（Kobuki底盘）" ], "tags": [ "导航与建图", "模仿学习" ], "limitations": "目标主要以图像指定，高层仍使用标准前沿探索；语言或坐标目标以及语义驱动探索尚未在本工作完成。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的统一导航与探索扩散策略研究，关注目标遮蔽和多模态动作。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2310.07896v1", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2310.07896v1", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/robodhruv/visualnav-transformer", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/robodhruv/visualnav-transformer", "category": "导航与建图 / 模仿学习", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "训练、微调、ROS 数据处理及部署脚本已发布。 列出公开数据；部分原训练数据未发布，需联系作者。 作者提供 GNM/ViNT/NoMaD 检查点链接。 作者明确训练混合了公开和未公开数据，不能宣称从零完全复现全部原始数据；不同方法共仓不应重复计数。 本次未运行训练、复现实验或实机控制。", "directions": [ "导航与建图", "模仿学习" ], "robotFilters": [ "LoCoBot（Kobuki底盘）" ], "original": { "id": "arxiv-2310.07896", "originalSourceUrl": "https://arxiv.org/pdf/2310.07896v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2310.07896v1", "finalDownloadUrl": "https://arxiv.org/pdf/2310.07896v1", "sha256": "05fe73918a7e454c6a006b7cba9aaec6bfc46ce96f65e28eb7a0967263a48af9", "bytes": 2364867, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:49.728730+00:00", "publicationDate": "2023/10/11", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2310.07896v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2310.07896v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "NoMaD: Goal Masked Diffusion Policies for Navigation and Exploration", "sourceVersion": "2310.07896v1", "sourceVersionDate": "2023/10/11", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2310.07896v1", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2310.07896" }, "originalAnalysis": { "id": "arxiv-2310.07896", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2310.07896v1", "sourceVersion": "2310.07896v1", "sectionsRead": [ "PDF第4页选段（提取文本L381–398）", "PDF第4页选段（提取文本L400–413）", "PDF第4–5页选段（提取文本L418–431）", "PDF第6页选段（提取文本L650–667）", "PDF第3页选段（提取文本L240–270）", "PDF第5页选段（提取文本L456–485）", "PDF第3–4页选段（提取文本L280–315）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "在ViNT式视觉编码中随机遮蔽目标token，用同一个动作扩散模型学习有目标导航与无目标探索，配合拓扑图进行长时序搜索。", "experimentsZh": "六种真实室内外环境，比较未知环境探索和已知环境图像目标导航；同时报告成功率与每次试验的碰撞数。", "resultsZh": "原文报告相比子目标图像扩散基线更高探索成功、参数量约少15倍；该对照是完整导航架构，不是任何扩散模型之间的通用结论。", "limitationsZh": "目标主要以图像指定，高层仍使用标准前沿探索；语言或坐标目标以及语义驱动探索尚未在本工作完成。", "reproductionZh": "固定目标遮蔽概率、动作扩散采样和拓扑图策略，保持训练数据与任务初态一致，并同时评估碰撞和成功。", "experimentType": "real", "robots": [ "LoCoBot（Kobuki底盘）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "在ViNT式视觉编码中随机遮蔽目标token，用同一个动作扩散模型学习有目标导航与无目标探索，配合拓扑图进行长时序搜索。" }, { "section": "原文实验与结果相关选段", "note": "六种真实室内外环境，比较未知环境探索和已知环境图像目标导航；同时报告成功率与每次试验的碰撞数。 原文报告相比子目标图像扩散基线更高探索成功、参数量约少15倍；该对照是完整导航架构，不是任何扩散模型之间的通用结论。" }, { "section": "局限与复现条件", "note": "目标主要以图像指定，高层仍使用标准前沿探索；语言或坐标目标以及语义驱动探索尚未在本工作完成。" } ], "originalSha256": "05fe73918a7e454c6a006b7cba9aaec6bfc46ce96f65e28eb7a0967263a48af9", "readingScopeSha256": "5966dac54b2762310e00ef0f822185f5b9b7baf968ce9dba9619110830ad17d2", "textExtractionSha256": "a858fc6944d58dabe2cd52a3b4bd59062a857d2e9d683ca9e938877daf755f9c" }, "experimentNote": "六种真实室内外环境，比较未知环境探索和已知环境图像目标导航；同时报告成功率与每次试验的碰撞数。", "contribution": "在ViNT式视觉编码中随机遮蔽目标token，用同一个动作扩散模型学习有目标导航与无目标探索，配合拓扑图进行长时序搜索。", "whyUseful": "固定目标遮蔽概率、动作扩散采样和拓扑图策略，保持训练数据与任务初态一致，并同时评估碰撞和成功。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "主代码 MIT；预训练权重和多来源数据仍需逐项查明条款。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2310.06114", "title": "Learning Interactive Real-World Simulators", "titleZh": "学习可交互的真实世界模拟器", "shortTitle": "UniSim", "date": "2023-10-09", "year": 2023, "url": "https://arxiv.org/abs/2310.06114", "paperUrl": "https://arxiv.org/abs/2310.06114", "projectUrl": "https://universal-simulator.github.io/", "codeUrl": null, "summary": "UniSim把语言、人类活动、机器人控制与相机运动统一为动作条件，使用视频U-Net扩散模型预测后续画面；多步交互通过最近观察递归生成。不同来源以T5嵌入和离散低层动作结合，模型没有显式可验证的刚体物理求解器。", "abstractZh": "UniSim把语言、人类活动、机器人控制与相机运动统一为动作条件，使用视频U-Net扩散模型预测后续画面；多步交互通过最近观察递归生成。不同来源以T5嵌入和离散低层动作结合，模型没有显式可验证的刚体物理求解器。\n既测Ego4D视频生成质量，也在Language Table研究两个用途：生成一万条长程轨迹并用事后目标重标训练高层策略；先行为克隆PaLI动作策略，再在生成器内用学习奖励和REINFORCE优化。论文展示真实桌面机器人执行，同时在生成环境中比较量化指标。", "category": "世界模型 / 生成式模拟器", "tags": [ "UniSim", "世界模型", "强化学习", "操作与抓取" ], "directions": [ "世界模型", "操作与抓取", "强化学习" ], "tier": "classic", "experimentType": "both", "experimentNote": "既测Ego4D视频生成质量，也在Language Table研究两个用途：生成一万条长程轨迹并用事后目标重标训练高层策略；先行为克隆PaLI动作策略，再在生成器内用学习奖励和REINFORCE优化。论文展示真实桌面机器人执行，同时在生成环境中比较量化指标。", "robots": [ "Language Table 桌面机器人（本次所读原文未明确型号）" ], "robotFilters": [], "robotNote": "已核实 Language Table 真实机器人实验；本文所读内容未明确部署机械臂具体型号，保留未知。", "codeStatus": "unknown", "status": "官方代码未核实", "trainingStatus": "官方代码未核实", "trainingNote": "论文提供模型与训练说明；本轮未核实完整官方训练仓库、权重或统一许可。", "codeStatusNote": "本轮未核实可复现该论文的官方训练仓库；不据此断言从未发布。", "license": "未核实", "contribution": "UniSim把语言、人类活动、机器人控制与相机运动统一为动作条件，使用视频U-Net扩散模型预测后续画面；多步交互通过最近观察递归生成。不同来源以T5嵌入和离散低层动作结合，模型没有显式可验证的刚体物理求解器。", "whyUseful": "5.6B生成模型训练使用512个TPU v3、20天，并依赖多套数据、逆动力学和奖励模型。复现应分开报告生成质量、模拟任务成功和真实闭环测试，核对数据与权重可获得性。", "limitations": "视觉真实不保证物理正确；有限历史可能遗忘遮挡状态，模型偏差会进入策略奖励。域标识改善域内生成却损害跨域迁移；模拟器本身使用真实机器人数据，不能称整个系统无需真实数据。", "caveats": "会产生幻觉、遗忘长期物体状态；域外形态泛化有限，不能建模不改变图像的力学状态。", "evidence": [ { "url": "https://arxiv.org/abs/2310.06114", "note": "首发日期与官方项目链接。" }, { "url": "https://arxiv.org/html/2310.06114v3", "note": "§4 与附录：Language Table 仿真和实机；§6 明确幻觉、记忆、泛化与纯视觉限制。" }, { "url": "https://arxiv.org/pdf/2310.06114v3", "note": "2.2 Architecture and Training：5.6B parameters; 512 TPU-v3 for 20 days." }, { "url": "https://arxiv.org/pdf/2310.06114v3", "note": "3.2 表1：Four recent frames improve FVD relative to one-frame conditioning; domain identifier hurts generalization." }, { "url": "https://arxiv.org/pdf/2310.06114v3", "note": "4.1 表2：RDG evaluated over five simulator runs; real robot examples separately described." }, { "url": "https://arxiv.org/pdf/2310.06114v3", "note": "4.2 表3：48 tasks; success assessed qualitatively from simulated video rollouts, 0.58→0.81." }, { "url": "https://arxiv.org/pdf/2310.06114v3", "note": "图7/8：Real Language Table robot deployment is shown; this text does not establish the manufacturer/model." } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2023：将生成式视频模型拓展为可用动作交互、用于训练策略的模拟器。", "freshness": "经典工作", "original": { "id": "arxiv-2310.06114", "originalSourceUrl": "https://arxiv.org/pdf/2310.06114v3", "versionedOriginalUrl": "https://arxiv.org/pdf/2310.06114v3", "finalDownloadUrl": "https://arxiv.org/pdf/2310.06114", "sha256": "956d21809f36cc107c82b3a32177ea555cacf032b3bde014fed1b958bc09c295", "bytes": 38316042, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:05.692358+00:00", "publicationDate": "2023/10/09", "publicationDateSourceUrl": "https://arxiv.org/abs/2310.06114", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2310.06114", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2310.06114", "pages": 25, "sourceIdentityStatus": "title_match", "sourceTitle": "Learning Interactive Real-World Simulators", "sourceVersion": "2310.06114v3", "sourceVersionDate": "2024/09/26", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.059507+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2310.06114" }, "originalAnalysis": { "id": "arxiv-2310.06114", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739386+00:00", "sourceUrl": "https://arxiv.org/pdf/2310.06114v3", "sectionsRead": [ "2.1–2.2 数据统一、观察预测与扩散训练", "3.1–3.2 长程模拟及消融表1", "4.1–4.3 机器人与视觉语言实验、表2–4" ], "methodsZh": "UniSim把语言、人类活动、机器人控制与相机运动统一为动作条件，使用视频U-Net扩散模型预测后续画面；多步交互通过最近观察递归生成。不同来源以T5嵌入和离散低层动作结合，模型没有显式可验证的刚体物理求解器。", "experimentsZh": "既测Ego4D视频生成质量，也在Language Table研究两个用途：生成一万条长程轨迹并用事后目标重标训练高层策略；先行为克隆PaLI动作策略，再在生成器内用学习奖励和REINFORCE优化。论文展示真实桌面机器人执行，同时在生成环境中比较量化指标。", "resultsZh": "作者报告最近四帧条件的FVD为211.3，单帧为315.69。五次模拟评测中，长程数据使全部积木目标距离改善指标从0.07升至0.34；48任务的模拟视频人工判定成功率由58%升至81%。这些数字不是实机成功率，真实迁移证据主要来自展示。", "limitationsZh": "视觉真实不保证物理正确；有限历史可能遗忘遮挡状态，模型偏差会进入策略奖励。域标识改善域内生成却损害跨域迁移；模拟器本身使用真实机器人数据，不能称整个系统无需真实数据。", "reproductionZh": "5.6B生成模型训练使用512个TPU v3、20天，并依赖多套数据、逆动力学和奖励模型。复现应分开报告生成质量、模拟任务成功和真实闭环测试，核对数据与权重可获得性。", "evidenceNotes": [ { "section/page": "2.2 Architecture and Training", "note": "5.6B parameters; 512 TPU-v3 for 20 days." }, { "section/page": "3.2 表1", "note": "Four recent frames improve FVD relative to one-frame conditioning; domain identifier hurts generalization." }, { "section/page": "4.1 表2", "note": "RDG evaluated over five simulator runs; real robot examples separately described." }, { "section/page": "4.2 表3", "note": "48 tasks; success assessed qualitatively from simulated video rollouts, 0.58→0.81." }, { "section/page": "图7/8", "note": "Real Language Table robot deployment is shown; this text does not establish the manufacturer/model." } ], "experimentType": "both", "robots": [ "Language Table 桌面机器人（本次所读原文未明确型号）" ], "corrections": [ "81%是生成模拟器内48任务的成功率，不是实机成功率。", "真实部署支持both，不能仅因名称simulator标作sim。" ], "sourceVersion": "2310.06114v3", "originalSha256": "956d21809f36cc107c82b3a32177ea555cacf032b3bde014fed1b958bc09c295" }, "analysisVerifiedAt": "2026-10-04T13:47:58.739386+00:00" }, { "id": "arxiv-2309.14341", "title": "Extreme Parkour with Legged Robots", "date": "2023-09-25", "url": "https://arxiv.org/abs/2309.14341v1", "titleZh": "腿式机器人的极限跑酷", "abstractZh": "用统一内积奖励和地形课程学习多样运动，再双重蒸馏特权策略与朝向估计，让深度图直接驱动关节动作和转向。\nA1配D435、Jetson NX，深度网络约10Hz、策略50Hz；仿真比较前进表现和踩边指标，实机展示爬高、跨跳和斜坡。", "summary": "用统一内积奖励和地形课程学习多样运动，再双重蒸馏特权策略与朝向估计，让深度图直接驱动关节动作和转向。", "experimentType": "both", "robots": [ "Unitree A1" ], "tags": [ "运动控制", "强化学习", "导航与建图" ], "limitations": "评测范围限制：结果来自指定A1、地形和深度管线，未建立一般安全保证，也未验证移动操作能力。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的端到端四足跑酷方案，用于比较奖励设计、朝向蒸馏和硬件条件。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "源码可用，存在使用限制", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2309.14341v1", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2309.14341v1", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/chengxuxin/extreme-parkour", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "source_available", "codeUrl": "https://github.com/chengxuxin/extreme-parkour", "category": "运动控制 / 强化学习", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "基础训练与相机蒸馏代码已发布。 训练现场生成仿真经验。 本次未核实官方现成检查点下载。 发布说明以仿真训练/回放为主；不能仅凭论文实机展示推定仓库已含完整上机部署和权重。 本次未运行训练、复现实验或实机控制。", "directions": [ "运动控制", "强化学习", "导航与建图" ], "robotFilters": [ "Unitree A1" ], "original": { "id": "arxiv-2309.14341", "originalSourceUrl": "https://arxiv.org/pdf/2309.14341v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2309.14341v1", "finalDownloadUrl": "https://arxiv.org/pdf/2309.14341v1", "sha256": "50f1cf2b73b9147111100994d1dfffbfd15a9a6c2412aee1921d4d3e58ee2285", "bytes": 6442374, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:27.282739+00:00", "publicationDate": "2023/09/25", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2309.14341v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2309.14341v1", "pages": 12, "sourceIdentityStatus": "title_match", "sourceTitle": "Extreme Parkour with Legged Robots", "sourceVersion": "2309.14341v1", "sourceVersionDate": "2023/09/25", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2309.14341v1", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2309.14341" }, "originalAnalysis": { "id": "arxiv-2309.14341", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2309.14341v1", "sourceVersion": "2309.14341v1", "sectionsRead": [ "PDF第3页选段（提取文本L210–227）", "PDF第7页选段（提取文本L488–505）", "PDF第8页选段（提取文本L599–612）", "PDF第10页选段（提取文本L775–807）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "用统一内积奖励和地形课程学习多样运动，再双重蒸馏特权策略与朝向估计，让深度图直接驱动关节动作和转向。", "experimentsZh": "A1配D435、Jetson NX，深度网络约10Hz、策略50Hz；仿真比较前进表现和踩边指标，实机展示爬高、跨跳和斜坡。", "resultsZh": "消融显示只奖励前进会诱导绕障或碰撞重试，朝向预测和抬脚项影响成功；原文训练预算为单3090少于20小时。", "limitationsZh": "评测范围限制：结果来自指定A1、地形和深度管线，未建立一般安全保证，也未验证移动操作能力。", "reproductionZh": "复现固定深度/本体时延、图像裁剪、地形课程和控制频率；跳跃峰值电流与结构载荷须另行核验。", "experimentType": "both", "robots": [ "Unitree A1" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "用统一内积奖励和地形课程学习多样运动，再双重蒸馏特权策略与朝向估计，让深度图直接驱动关节动作和转向。" }, { "section": "原文实验与结果相关选段", "note": "A1配D435、Jetson NX，深度网络约10Hz、策略50Hz；仿真比较前进表现和踩边指标，实机展示爬高、跨跳和斜坡。 消融显示只奖励前进会诱导绕障或碰撞重试，朝向预测和抬脚项影响成功；原文训练预算为单3090少于20小时。" }, { "section": "局限与复现条件", "note": "评测范围限制：结果来自指定A1、地形和深度管线，未建立一般安全保证，也未验证移动操作能力。" } ], "originalSha256": "50f1cf2b73b9147111100994d1dfffbfd15a9a6c2412aee1921d4d3e58ee2285", "readingScopeSha256": "b08c216ec819c18fd3abe8e7c970687dba698f286c749fd9ddfa27eae4d304e1", "textExtractionSha256": "ed2bb57e72ec7602ed1156d25b2e711b2cb34454c7328e26acc30354743db5b8" }, "experimentNote": "A1配D435、Jetson NX，深度网络约10Hz、策略50Hz；仿真比较前进表现和踩边指标，实机展示爬高、跨跳和斜坡。", "contribution": "用统一内积奖励和地形课程学习多样运动，再双重蒸馏特权策略与朝向估计，让深度图直接驱动关节动作和转向。", "whyUseful": "复现固定深度/本体时延、图像裁剪、地形课程和控制频率；跳跃峰值电流与结构载荷须另行核验。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "主许可证 CC BY-NC 4.0，非商业；继承代码和 Isaac Gym 另有许可。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2309.05665", "title": "Robot Parkour Learning", "date": "2023-09-11", "url": "https://arxiv.org/abs/2309.05665v2", "titleZh": "机器人跑酷学习", "abstractZh": "先在允许障碍穿透的软动力学课程中探索技能，再恢复真实约束微调；用DAgger把爬越、跨跳、匍匐、侧倾及奔跑蒸馏为深度视觉策略。\nIsaac Gym训练后部署到A1和Go1，使用Jetson NX及RealSense D435；仿真按三个种子、每技能100次测试比较完成率和距离。", "summary": "先在允许障碍穿透的软动力学课程中探索技能，再恢复真实约束微调；用DAgger把爬越、跨跳、匍匐、侧倾及奔跑蒸馏为深度视觉策略。", "experimentType": "both", "robots": [ "Unitree A1", "Unitree Go1" ], "tags": [ "运动控制", "强化学习", "导航与建图" ], "limitations": "新技能依赖手工搭建新训练地形；深度预处理、视觉时延标定和电机保护对仿真迁移很关键。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的视觉四足跑酷学习方案，关注软动力学课程和技能蒸馏。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2309.05665v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2309.05665v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/ZiwenZhuang/parkour", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/ZiwenZhuang/parkour", "category": "运动控制 / 强化学习", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "A1/Go2 仿真训练和 Go1 配置已发布。 训练中生成仿真经验；本次未核实独立数据集。 可训练后导出；本次未核实统一预训练检查点。 A1 上机代码在 README 仍未勾选；Go1/Go2 配置不保证论文同等性能，不能把所有型号都标成部署完备。 本次未运行训练、复现实验或实机控制。", "directions": [ "运动控制", "强化学习", "导航与建图" ], "robotFilters": [ "Unitree A1", "Unitree Go1" ], "original": { "id": "arxiv-2309.05665", "originalSourceUrl": "https://arxiv.org/pdf/2309.05665v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2309.05665v2", "finalDownloadUrl": "https://arxiv.org/pdf/2309.05665v2", "sha256": "373da1078eaafff39404d6d3533f8718367ad8b1cca55125128e395692f99b69", "bytes": 7511090, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:19.923951+00:00", "publicationDate": "2023/09/11", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2309.05665v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2309.05665v2", "pages": 20, "sourceIdentityStatus": "title_match", "sourceTitle": "Robot Parkour Learning", "sourceVersion": "2309.05665v2", "sourceVersionDate": "2023/09/12", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2309.05665v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2309.05665" }, "originalAnalysis": { "id": "arxiv-2309.05665", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2309.05665v2", "sourceVersion": "2309.05665v2", "sectionsRead": [ "PDF第3–4页选段（提取文本L142–159）", "PDF第6页选段（提取文本L284–297）", "PDF第6页选段（提取文本L390–407）", "PDF第6–7页选段（提取文本L471–484）", "PDF第8页选段（提取文本L586–603）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "先在允许障碍穿透的软动力学课程中探索技能，再恢复真实约束微调；用DAgger把爬越、跨跳、匍匐、侧倾及奔跑蒸馏为深度视觉策略。", "experimentsZh": "Isaac Gym训练后部署到A1和Go1，使用Jetson NX及RealSense D435；仿真按三个种子、每技能100次测试比较完成率和距离。", "resultsZh": "单一视觉策略可根据障碍执行多种跑酷技能；两阶段训练降低困难探索成本，但现场展示不等于无保护条件下可保证成功。", "limitationsZh": "新技能依赖手工搭建新训练地形；深度预处理、视觉时延标定和电机保护对仿真迁移很关键。", "reproductionZh": "固定障碍分布、软约束课程及DAgger教师切换，记录控制延迟和电机安全限值，先用保护设施验证。", "experimentType": "both", "robots": [ "Unitree A1", "Unitree Go1" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "先在允许障碍穿透的软动力学课程中探索技能，再恢复真实约束微调；用DAgger把爬越、跨跳、匍匐、侧倾及奔跑蒸馏为深度视觉策略。" }, { "section": "原文实验与结果相关选段", "note": "Isaac Gym训练后部署到A1和Go1，使用Jetson NX及RealSense D435；仿真按三个种子、每技能100次测试比较完成率和距离。 单一视觉策略可根据障碍执行多种跑酷技能；两阶段训练降低困难探索成本，但现场展示不等于无保护条件下可保证成功。" }, { "section": "局限与复现条件", "note": "新技能依赖手工搭建新训练地形；深度预处理、视觉时延标定和电机保护对仿真迁移很关键。" } ], "originalSha256": "373da1078eaafff39404d6d3533f8718367ad8b1cca55125128e395692f99b69", "readingScopeSha256": "0549287915420fae8a2e347d9dff81f1de02a502ee0a0da2f9eee3f6df3e9d61", "textExtractionSha256": "5c1fbcc1e736914bfe5c785c2d967ac3549ce9733ff1c8f54c81796b3f597368" }, "experimentNote": "Isaac Gym训练后部署到A1和Go1，使用Jetson NX及RealSense D435；仿真按三个种子、每技能100次测试比较完成率和距离。", "contribution": "先在允许障碍穿透的软动力学课程中探索技能，再恢复真实约束微调；用DAgger把爬越、跨跳、匍匐、侧倾及奔跑蒸馏为深度视觉策略。", "whyUseful": "固定障碍分布、软约束课程及DAgger教师切换，记录控制延迟和电机安全限值，先用保护设施验证。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "主仓库 MIT；Isaac Gym、Unitree SDK 与第三方依赖分别核对。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2308.12952", "title": "BridgeData V2: A Dataset for Robot Learning at Scale", "date": "2023-08-24", "url": "https://arxiv.org/abs/2308.12952v3", "titleZh": "BridgeData V2：用于规模化机器人学习的数据集", "abstractZh": "在多种布景、物体和相机位置中采集任务混合轨迹，定期随机化环境，并提供目标图像或语言条件以支持多类离线学习算法。\n60,096条轨迹；以固定肩后视角训练行为克隆、扩散、ACT、RT-1等方法，使用六维末端增量和夹爪开合动作，在真实任务中比较。", "summary": "在多种布景、物体和相机位置中采集任务混合轨迹，定期随机化环境，并提供目标图像或语言条件以支持多类离线学习算法。", "experimentType": "real", "robots": [ "Trossen WidowX 250" ], "tags": [ "数据集与基准", "操作与抓取", "模仿学习" ], "limitations": "以低成本单臂为主，强力、动态、精密插入任务覆盖不足；数据全部来自单一机构，不代表跨形态数据集。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "foundation", "tierNote": "编辑分类：广泛使用的多任务机器人数据基础设施，作为数据格式、目标条件学习和离线训练入口。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2308.12952v3", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2308.12952v3", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/rail-berkeley/bridge_data_v2", "note": "已核验官方代码与文档；未运行复现；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/rail-berkeley/bridge_data_v2", "category": "数据集与基准 / 操作与抓取", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "README已核实数据预处理、BC/RL训练与WidowX控制栈入口；这份JAX实现不等于论文所有基线都在同一仓库。 本次未运行训练、复现实验或实机控制。", "directions": [ "数据集与基准", "操作与抓取", "模仿学习" ], "robotFilters": [ "Trossen WidowX 250" ], "original": { "id": "arxiv-2308.12952", "originalSourceUrl": "https://arxiv.org/pdf/2308.12952v3", "versionedOriginalUrl": "https://arxiv.org/pdf/2308.12952v3", "finalDownloadUrl": "https://arxiv.org/pdf/2308.12952v3", "sha256": "aa5fdac48bf9122b22f88150d45575cb0685f2a2f761ca6fc570be743022819b", "bytes": 3258884, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T14:59:35.274215+00:00", "publicationDate": "2023/08/24", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2308.12952v3", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2308.12952v3", "pages": 15, "sourceIdentityStatus": "title_match", "sourceTitle": "BridgeData V2: A Dataset for Robot Learning at Scale", "sourceVersion": "2308.12952v3", "sourceVersionDate": "2024/01/17", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2308.12952v3", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2308.12952" }, "originalAnalysis": { "id": "arxiv-2308.12952", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2308.12952v3", "sourceVersion": "2308.12952v3", "sectionsRead": [ "PDF第4页选段（提取文本L295–308）", "PDF第4页选段（提取文本L316–333）", "PDF第5–6页选段（提取文本L386–403）", "PDF第6–7页选段（提取文本L442–455）", "PDF第8页选段（提取文本L751–768）", "PDF第9页选段（提取文本L825–845）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "在多种布景、物体和相机位置中采集任务混合轨迹，定期随机化环境，并提供目标图像或语言条件以支持多类离线学习算法。", "experimentsZh": "60,096条轨迹；以固定肩后视角训练行为克隆、扩散、ACT、RT-1等方法，使用六维末端增量和夹爪开合动作，在真实任务中比较。", "resultsZh": "原文每任务测试10次，RT-1优于所用语言条件BC基线，目标条件方法表现接近；跨物体和环境结果依赖方法及任务。", "limitationsZh": "以低成本单臂为主，强力、动态、精密插入任务覆盖不足；数据全部来自单一机构，不代表跨形态数据集。", "reproductionZh": "固定数据版本、图像尺寸和任务划分；区分原始多相机数据与论文仅肩后视角的训练设定。", "experimentType": "real", "robots": [ "Trossen WidowX 250" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "在多种布景、物体和相机位置中采集任务混合轨迹，定期随机化环境，并提供目标图像或语言条件以支持多类离线学习算法。" }, { "section": "原文实验与结果相关选段", "note": "60,096条轨迹；以固定肩后视角训练行为克隆、扩散、ACT、RT-1等方法，使用六维末端增量和夹爪开合动作，在真实任务中比较。 原文每任务测试10次，RT-1优于所用语言条件BC基线，目标条件方法表现接近；跨物体和环境结果依赖方法及任务。" }, { "section": "局限与复现条件", "note": "以低成本单臂为主，强力、动态、精密插入任务覆盖不足；数据全部来自单一机构，不代表跨形态数据集。" } ], "originalSha256": "aa5fdac48bf9122b22f88150d45575cb0685f2a2f761ca6fc570be743022819b", "readingScopeSha256": "e6e6595147184bf8426b7057e17636a1abc5966a47b10cb9e2edc6a583184af7", "textExtractionSha256": "8423997e2cc7efeb0674d3537a4adc51fc96522b8408465dbd8036b3ee463864" }, "experimentNote": "60,096条轨迹；以固定肩后视角训练行为克隆、扩散、ACT、RT-1等方法，使用六维末端增量和夹爪开合动作，在真实任务中比较。", "contribution": "在多种布景、物体和相机位置中采集任务混合轨迹，定期随机化环境，并提供目标图像或语言条件以支持多类离线学习算法。", "whyUseful": "固定数据版本、图像尺寸和任务划分；区分原始多相机数据与论文仅肩后视角的训练设定。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "MIT；Bridge数据单独核对许可。", "codeReleaseScope": "已核验官方代码与文档；未运行复现" }, { "id": "arxiv-2307.15818", "title": "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control", "titleZh": "RT-2：将网络知识迁移到机器人控制的视觉－语言－动作模型", "shortTitle": "RT-2", "date": "2023-07-28", "year": 2023, "url": "https://arxiv.org/abs/2307.15818", "codeUrl": null, "category": "视觉语言动作", "tags": [ "VLA", "网络知识迁移", "动作词元", "实机" ], "tier": "foundation", "summary": "RT-2将平移、旋转、夹爪及停止动作编码为离散文本token，以PaLI-X或PaLM-E联合微调机器人示范和原互联网视觉语言任务；动作模式限制合法输出词表，保留语义先验并直接生成闭环控制命令。", "abstractZh": "RT-2将平移、旋转、夹爪及停止动作编码为离散文本token，以PaLI-X或PaLM-E联合微调机器人示范和原互联网视觉语言任务；动作模式限制合法输出词表，保留语义先验并直接生成闭环控制命令。\n主实验使用七自由度移动操作机器人，约6000条评估轨迹，分已见任务、未见物体/背景/环境、符号与推理等设置。机器人数据来自13台设备17个月采集；另以小型PaLI3B在Language-Table仿真评估并展示真实推物。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "Google七自由度移动操作机器人" ], "codeStatus": "unknown", "status": "已核官方论文与项目页；未确认可公开使用的官方 RT-2 实现或权重。", "trainingNote": "原文包含基于 PaLI-X 与 PaLM-E 的 RT-2 变体；大模型通过多 TPU 云端推理。不要把 RT-1 仓库当作 RT-2。", "whyUseful": "复现需明确PaLI-X/PaLM-E版本、机器人与网页共训比例、动作分箱及任务难度，分开主移动机器人和Language-Table分支。主模型与训练资源存在可获取限制，不能把相关开源基准视作RT-2完整公开实现。", "contribution": "RT-2将平移、旋转、夹爪及停止动作编码为离散文本token，以PaLI-X或PaLM-E联合微调机器人示范和原互联网视觉语言任务；动作模式限制合法输出词表，保留语义先验并直接生成闭环控制命令。", "limitations": "网络知识没有自动创造新的物理动作，运动能力仍受机器人数据分布限制。55B模型依赖多TPU云端、约1–3Hz，5B约5Hz，难满足更高频控制；研究不证明长时域安全或任意环境可靠性。", "caveats": "代码状态记为未确认，而非仅因未找到链接就判定闭源。", "evidence": [ { "url": "https://arxiv.org/abs/2307.15818", "note": "首发日期与 VLA 定义。" }, { "url": "https://arxiv.org/html/2307.15818v1", "note": "第 3–4 节模型变体、动作编码、云端推理和实机评测。" }, { "url": "https://robotics-transformer2.github.io/", "note": "作者项目页的模型与实验说明。" }, { "url": "https://arxiv.org/pdf/2307.15818", "note": "§3.3：55B使用网络访问多TPU服务，1–3Hz。" }, { "url": "https://arxiv.org/pdf/2307.15818", "note": "Appendix H Table 4：未见均值62%对RT-1 32%；已见约持平。" }, { "url": "https://arxiv.org/pdf/2307.15818", "note": "§5：语义泛化不产生训练数据之外新物理技能。" } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "视觉语言动作" ], "robotFilters": [ "Google Robot" ], "verificationNote": "已核论文与作者项目页；代码公开性未获充分一手证据。", "original": { "id": "arxiv-2307.15818", "originalSourceUrl": "https://arxiv.org/pdf/2307.15818v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2307.15818v1", "finalDownloadUrl": "https://arxiv.org/pdf/2307.15818", "sha256": "0a62dc36bbbfdea232ad45a0ab75c82e9cf535b55333ff22c02d5da1897ccac3", "bytes": 15668163, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:09.098735+00:00", "publicationDate": "2023/07/28", "publicationDateSourceUrl": "https://arxiv.org/abs/2307.15818", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2307.15818", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2307.15818", "pages": 26, "sourceIdentityStatus": "title_match", "sourceTitle": "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control", "sourceVersion": "2307.15818v1", "sourceVersionDate": "2023/07/28", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.072016+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2307.15818" }, "originalAnalysis": { "id": "arxiv-2307.15818", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090137+00:00", "sourceUrl": "https://arxiv.org/pdf/2307.15818", "sectionsRead": [ "§3.2–3.3", "§4.1–4.4", "§5", "Appendix H Table 4–5" ], "methodsZh": "RT-2将平移、旋转、夹爪及停止动作编码为离散文本token，以PaLI-X或PaLM-E联合微调机器人示范和原互联网视觉语言任务；动作模式限制合法输出词表，保留语义先验并直接生成闭环控制命令。", "experimentsZh": "主实验使用七自由度移动操作机器人，约6000条评估轨迹，分已见任务、未见物体/背景/环境、符号与推理等设置。机器人数据来自13台设备17个月采集；另以小型PaLI3B在Language-Table仿真评估并展示真实推物。", "resultsZh": "作者附表4显示RT-2两版本未见设置均62%，RT-1为32%、MOO35%；已见任务约91–93%，RT-1为92%。因此优势主要在指定泛化轴，不能说所有任务翻倍；新语义可重新调用已有动作。", "limitationsZh": "网络知识没有自动创造新的物理动作，运动能力仍受机器人数据分布限制。55B模型依赖多TPU云端、约1–3Hz，5B约5Hz，难满足更高频控制；研究不证明长时域安全或任意环境可靠性。", "reproductionZh": "复现需明确PaLI-X/PaLM-E版本、机器人与网页共训比例、动作分箱及任务难度，分开主移动机器人和Language-Table分支。主模型与训练资源存在可获取限制，不能把相关开源基准视作RT-2完整公开实现。", "experimentType": "both", "robots": [ "Google七自由度移动操作机器人" ], "evidenceNotes": [ { "section/page": "§3.3", "note": "55B使用网络访问多TPU服务，1–3Hz。" }, { "section/page": "Appendix H Table 4", "note": "未见均值62%对RT-1 32%；已见约持平。" }, { "section/page": "§5", "note": "语义泛化不产生训练数据之外新物理技能。" } ] }, "experimentNote": "主实验使用七自由度移动操作机器人，约6000条评估轨迹，分已见任务、未见物体/背景/环境、符号与推理等设置。机器人数据来自13台设备17个月采集；另以小型PaLI3B在Language-Table仿真评估并展示真实推物。", "analysisVerifiedAt": "2026-10-04T13:50:35.090137+00:00" }, { "id": "arxiv-2307.05973", "title": "VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models", "date": "2023-07-12", "url": "https://arxiv.org/abs/2307.05973v2", "titleZh": "VoxPoser：用语言模型构建可组合三维价值图进行机器人操作", "abstractZh": "LLM与视觉语言感知生成三维可供性和约束价值图，再用模型式运动优化合成闭环末端轨迹；接触任务可结合在线动力学学习。\n真实桌面操作与13类仿真任务，区分已见/未见指令及物体属性，并分析感知、动力学及任务规格错误。", "summary": "LLM与视觉语言感知生成三维可供性和约束价值图，再用模型式运动优化合成闭环末端轨迹；接触任务可结合在线动力学学习。", "experimentType": "both", "robots": [ "Franka Emika Panda" ], "tags": [ "操作与抓取", "视觉语言动作" ], "limitations": "依赖外部感知、人工提示词和动力学；仅规划末端轨迹，不能自动保证整个机械臂避障。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的语言模型、三维价值图与运动规划组合研究。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2307.05973v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2307.05973v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/huangwl18/VoxPoser", "note": "已发布仿真 demo；真实感知和部署适配不完整；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/huangwl18/VoxPoser", "category": "操作与抓取 / 视觉语言动作", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "方法为零样本组合，不提供策略训练配方。 无需任务训练集；仿真使用 RLBench 提供的物体掩码。 不发布自有统一策略权重；依赖外部模型/API。 真实实验的物体名称到分割掩码感知流水线没有包含在仓库中；生成代码需隔离检查，不能直接驱动实机。 本次未运行训练、复现实验或实机控制。", "directions": [ "操作与抓取", "视觉语言动作" ], "robotFilters": [ "Franka Emika Panda" ], "original": { "id": "arxiv-2307.05973", "originalSourceUrl": "https://arxiv.org/pdf/2307.05973v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2307.05973v2", "finalDownloadUrl": "https://arxiv.org/pdf/2307.05973v2", "sha256": "cf1cadcf7a1ee44979908c07698f87755c9b5340a79420db509a92d6d2533d42", "bytes": 9374922, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T14:59:45.493022+00:00", "publicationDate": "2023/07/12", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2307.05973v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2307.05973v2", "pages": 23, "sourceIdentityStatus": "title_match", "sourceTitle": "VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models", "sourceVersion": "2307.05973v2", "sourceVersionDate": "2023/11/02", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2307.05973v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2307.05973" }, "originalAnalysis": { "id": "arxiv-2307.05973", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2307.05973v2", "sourceVersion": "2307.05973v2", "sectionsRead": [ "PDF第4页选段（提取文本L252–269）", "PDF第7页选段（提取文本L519–532）", "PDF第8页选段（提取文本L603–620）", "PDF第3–4页选段（提取文本L172–218）", "PDF第8–9页选段（提取文本L617–640）", "PDF第19–20页选段（提取文本L1092–1124）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "LLM与视觉语言感知生成三维可供性和约束价值图，再用模型式运动优化合成闭环末端轨迹；接触任务可结合在线动力学学习。", "experimentsZh": "真实桌面操作与13类仿真任务，区分已见/未见指令及物体属性，并分析感知、动力学及任务规格错误。", "resultsZh": "价值图规划比文中的动作原语基线更能处理扰动；错误分解显示感知质量和动力学模型仍是主要瓶颈。", "limitationsZh": "依赖外部感知、人工提示词和动力学；仅规划末端轨迹，不能自动保证整个机械臂避障。", "reproductionZh": "固定LLM、检测器、分割器、提示词和坐标标定，记录重规划开销；LLM生成约束应在独立安全层再次验证。", "experimentType": "both", "robots": [ "Franka Emika Panda" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "LLM与视觉语言感知生成三维可供性和约束价值图，再用模型式运动优化合成闭环末端轨迹；接触任务可结合在线动力学学习。" }, { "section": "原文实验与结果相关选段", "note": "真实桌面操作与13类仿真任务，区分已见/未见指令及物体属性，并分析感知、动力学及任务规格错误。 价值图规划比文中的动作原语基线更能处理扰动；错误分解显示感知质量和动力学模型仍是主要瓶颈。" }, { "section": "局限与复现条件", "note": "依赖外部感知、人工提示词和动力学；仅规划末端轨迹，不能自动保证整个机械臂避障。" } ], "originalSha256": "cf1cadcf7a1ee44979908c07698f87755c9b5340a79420db509a92d6d2533d42", "readingScopeSha256": "0fed6ccb9963b375210bb203895d098d057988401018f15115fc40fe3d3c5c74", "textExtractionSha256": "a3e0eebce0ed99fac509b5c8fe1108e504e1a8b1d66293831a7073ac9f439fe7" }, "experimentNote": "真实桌面操作与13类仿真任务，区分已见/未见指令及物体属性，并分析感知、动力学及任务规格错误。", "contribution": "LLM与视觉语言感知生成三维可供性和约束价值图，再用模型式运动优化合成闭环末端轨迹；接触任务可结合在线动力学学习。", "whyUseful": "固定LLM、检测器、分割器、提示词和坐标标定，记录重规划开销；LLM生成约束应在独立安全层再次验证。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "仓库 MIT；外部 LLM 服务、RLBench 和 CoppeliaSim 分别适用各自条款。", "codeReleaseScope": "已发布仿真 demo；真实感知和部署适配不完整" }, { "id": "arxiv-2306.14846", "title": "ViNT: A Foundation Model for Visual Navigation", "date": "2023-06-26", "url": "https://arxiv.org/abs/2306.14846v2", "titleZh": "ViNT：视觉导航基础模型", "abstractZh": "用Transformer处理历史观测与目标图像，预测归一化相对航点和到达时间；跨机器人数据训练后，以扩散子目标和拓扑搜索扩展到长距离。\n使用八种平台逾100小时真实轨迹预训练，在五种机器人上测试零样本迁移、探索、微调和新目标模态适配。", "summary": "用Transformer处理历史观测与目标图像，预测归一化相对航点和到达时间；跨机器人数据训练后，以扩散子目标和拓扑搜索扩展到长距离。", "experimentType": "real", "robots": [ "Clearpath Jackal", "LoCoBot（Kobuki底盘）", "Vizbot（Roomba底盘）", "Unitree Go1", "无人机（原文未细分型号）" ], "tags": [ "导航与建图", "通用机器人策略", "模仿学习" ], "limitations": "假设动作结构相似，不能控制无人机高度或直接接入LiDAR；Transformer推理负担对受限电源平台仍重要。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的跨机器人视觉导航预训练研究；论文称基础模型不等于本目录基础设施分类。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2306.14846v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2306.14846v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/robodhruv/visualnav-transformer", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/robodhruv/visualnav-transformer", "category": "导航与建图 / 通用机器人策略", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "训练、微调、ROS 数据处理及部署脚本已发布。 列出公开数据；部分原训练数据未发布，需联系作者。 作者提供 GNM/ViNT/NoMaD 检查点链接。 作者明确训练混合了公开和未公开数据，不能宣称从零完全复现全部原始数据；不同方法共仓不应重复计数。 本次未运行训练、复现实验或实机控制。", "directions": [ "导航与建图", "通用机器人策略", "模仿学习" ], "robotFilters": [ "Clearpath Jackal", "LoCoBot（Kobuki底盘）", "Vizbot（Roomba底盘）", "Unitree Go1", "无人机（原文未细分型号）" ], "original": { "id": "arxiv-2306.14846", "originalSourceUrl": "https://arxiv.org/pdf/2306.14846v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2306.14846v2", "finalDownloadUrl": "https://arxiv.org/pdf/2306.14846v2", "sha256": "175369795be4c276d74c76aa39c4c977b742d994bb717fcdf582052dc311d37d", "bytes": 7939854, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:53.519926+00:00", "publicationDate": "2023/06/26", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2306.14846v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2306.14846v2", "pages": 25, "sourceIdentityStatus": "title_match", "sourceTitle": "ViNT: A Foundation Model for Visual Navigation", "sourceVersion": "2306.14846v2", "sourceVersionDate": "2023/10/24", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2306.14846v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2306.14846" }, "originalAnalysis": { "id": "arxiv-2306.14846", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2306.14846v2", "sourceVersion": "2306.14846v2", "sectionsRead": [ "PDF第6–7页选段（提取文本L348–365）", "PDF第7页选段（提取文本L395–408）", "PDF第7页选段（提取文本L419–436）", "PDF第11页选段（提取文本L722–739）", "PDF第3–4页选段（提取文本L188–225）", "PDF第22页选段（提取文本L1360–1390）", "PDF第22页选段（提取文本L1422–1449）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "用Transformer处理历史观测与目标图像，预测归一化相对航点和到达时间；跨机器人数据训练后，以扩散子目标和拓扑搜索扩展到长距离。", "experimentsZh": "使用八种平台逾100小时真实轨迹预训练，在五种机器人上测试零样本迁移、探索、微调和新目标模态适配。", "resultsZh": "异构数据有利于跨机器人迁移；每个平台仍需要自己的低层航点控制器，长距离性能还依赖高层启发式搜索。", "limitationsZh": "假设动作结构相似，不能控制无人机高度或直接接入LiDAR；Transformer推理负担对受限电源平台仍重要。", "reproductionZh": "复现航点速度归一化、相机与历史窗口、目标token及低层控制；无人机平面导航不能表述成任意三维飞行能力。", "experimentType": "real", "robots": [ "Clearpath Jackal", "LoCoBot（Kobuki底盘）", "Vizbot（Roomba底盘）", "Unitree Go1", "无人机（原文未细分型号）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "用Transformer处理历史观测与目标图像，预测归一化相对航点和到达时间；跨机器人数据训练后，以扩散子目标和拓扑搜索扩展到长距离。" }, { "section": "原文实验与结果相关选段", "note": "使用八种平台逾100小时真实轨迹预训练，在五种机器人上测试零样本迁移、探索、微调和新目标模态适配。 异构数据有利于跨机器人迁移；每个平台仍需要自己的低层航点控制器，长距离性能还依赖高层启发式搜索。" }, { "section": "局限与复现条件", "note": "假设动作结构相似，不能控制无人机高度或直接接入LiDAR；Transformer推理负担对受限电源平台仍重要。" } ], "originalSha256": "175369795be4c276d74c76aa39c4c977b742d994bb717fcdf582052dc311d37d", "readingScopeSha256": "df45db5f79cfacf2c8301cb11dcb5e366b879b4016dc2f184b4e2b6508612aae", "textExtractionSha256": "eb5214f67edb8badb3f9a03d8bbf45869309ef2b53e434d269cfcda304b6ce91" }, "experimentNote": "使用八种平台逾100小时真实轨迹预训练，在五种机器人上测试零样本迁移、探索、微调和新目标模态适配。", "contribution": "用Transformer处理历史观测与目标图像，预测归一化相对航点和到达时间；跨机器人数据训练后，以扩散子目标和拓扑搜索扩展到长距离。", "whyUseful": "复现航点速度归一化、相机与历史窗口、目标token及低层控制；无人机平面导航不能表述成任意三维飞行能力。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "主代码 MIT；预训练权重和多来源数据仍需逐项查明条款。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2306.14896", "title": "RVT: Robotic View Transformer for 3D Object Manipulation", "date": "2023-06-26", "url": "https://arxiv.org/abs/2306.14896v1", "titleZh": "RVT：用于三维物体操作的机器人视图Transformer", "abstractZh": "先把RGB-D重建为点云，再从机器人基座周围虚拟视角重渲染；多视图Transformer融合语言并把预测反投影到三维末端动作。\n18项RLBench仿真任务，与PerAct在相同GPU类型下比较；真实Panda配Azure Kinect，五类任务共采集51条示教。", "summary": "先把RGB-D重建为点云，再从机器人基座周围虚拟视角重渲染；多视图Transformer融合语言并把预测反投影到三维末端动作。", "experimentType": "both", "robots": [ "Franka Panda" ], "tags": [ "操作与抓取", "模仿学习", "视觉语言动作" ], "limitations": "视角选择仍人为设计，需要机器人到相机的外参；低层规划器和关键帧控制限制高频接触操作。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的多视图三维操作方法；与PerAct对照有教学价值，不将模型名称当成基础设施等级。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "源码可用，存在使用限制", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2306.14896v1", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2306.14896v1", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/NVlabs/RVT", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "source_available", "codeUrl": "https://github.com/NVlabs/RVT", "category": "操作与抓取 / 模仿学习", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "RVT 与 RVT-2 训练脚本和配置均已发布。 使用 PerAct 的 RLBench 数据；另提供预生成回放缓存。 Hugging Face 提供 RVT 与 RVT-2 权重。 不要把 RVT-2 配置、权重或性能直接归到原始 RVT；评估随机性和最终检查点选择会影响比较。 本次未运行训练、复现实验或实机控制。", "directions": [ "操作与抓取", "模仿学习", "视觉语言动作" ], "robotFilters": [ "Franka Panda" ], "original": { "id": "arxiv-2306.14896", "originalSourceUrl": "https://arxiv.org/pdf/2306.14896v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2306.14896v1", "finalDownloadUrl": "https://arxiv.org/pdf/2306.14896v1", "sha256": "73f97253e2252b274ca6f96a14279685cd7c5068575e3bb3679646d0d71707fe", "bytes": 6045133, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:01:25.339834+00:00", "publicationDate": "2023/06/26", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2306.14896v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2306.14896v1", "pages": 16, "sourceIdentityStatus": "title_match", "sourceTitle": "RVT: Robotic View Transformer for 3D Object Manipulation", "sourceVersion": "2306.14896v1", "sourceVersionDate": "2023/06/26", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2306.14896v1", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2306.14896" }, "originalAnalysis": { "id": "arxiv-2306.14896", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2306.14896v1", "sourceVersion": "2306.14896v1", "sectionsRead": [ "PDF第5页选段（提取文本L342–355）", "PDF第5页选段（提取文本L370–383）", "PDF第6页选段（提取文本L402–419）", "PDF第8–9页选段（提取文本L760–777）", "PDF第3–4页选段（提取文本L158–195）", "PDF第7–8页选段（提取文本L715–750）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "先把RGB-D重建为点云，再从机器人基座周围虚拟视角重渲染；多视图Transformer融合语言并把预测反投影到三维末端动作。", "experimentsZh": "18项RLBench仿真任务，与PerAct在相同GPU类型下比较；真实Panda配Azure Kinect，五类任务共采集51条示教。", "resultsZh": "论文报告在18项任务中16项优于PerAct，推理约快2.3倍；真实堆块与按洗手液任务分别达到100%及80%，仅限文中测试。", "limitationsZh": "视角选择仍人为设计，需要机器人到相机的外参；低层规划器和关键帧控制限制高频接触操作。", "reproductionZh": "锁定原版RVT配置，别把仓库后续RVT-2结果当成本论文；保持评估变体、随机规划重复次数和GPU预算一致。", "experimentType": "both", "robots": [ "Franka Panda" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "先把RGB-D重建为点云，再从机器人基座周围虚拟视角重渲染；多视图Transformer融合语言并把预测反投影到三维末端动作。" }, { "section": "原文实验与结果相关选段", "note": "18项RLBench仿真任务，与PerAct在相同GPU类型下比较；真实Panda配Azure Kinect，五类任务共采集51条示教。 论文报告在18项任务中16项优于PerAct，推理约快2.3倍；真实堆块与按洗手液任务分别达到100%及80%，仅限文中测试。" }, { "section": "局限与复现条件", "note": "视角选择仍人为设计，需要机器人到相机的外参；低层规划器和关键帧控制限制高频接触操作。" } ], "originalSha256": "73f97253e2252b274ca6f96a14279685cd7c5068575e3bb3679646d0d71707fe", "readingScopeSha256": "da1bf30eb5edc1637daf2f2a5d4d1dfd4e297ba66cf2fbac9b5222832235c67b", "textExtractionSha256": "05612ae92aaf3c2c19314ff5b4acbe5e2dcd6e4e96c1f4db9ccecbcb6e87bd9a" }, "experimentNote": "18项RLBench仿真任务，与PerAct在相同GPU类型下比较；真实Panda配Azure Kinect，五类任务共采集51条示教。", "contribution": "先把RGB-D重建为点云，再从机器人基座周围虚拟视角重渲染；多视图Transformer融合语言并把预测反投影到三维末端动作。", "whyUseful": "锁定原版RVT配置，别把仓库后续RVT-2结果当成本论文；保持评估变体、随机规划重复次数和GPU预算一致。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "代码 NVIDIA Source Code License，含非商业限制；预训练模型 CC BY-NC-SA 4.0。RLBench 等另有许可。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2306.11706", "title": "RoboCat: A Self-Improving Generalist Agent for Robotic Manipulation", "titleZh": "RoboCat：可自我改进的通用机器人操作智能体", "shortTitle": "RoboCat", "date": "2023-06-20", "year": 2023, "url": "https://arxiv.org/abs/2306.11706", "paperUrl": "https://arxiv.org/abs/2306.11706", "projectUrl": null, "codeUrl": null, "summary": "把图像经冻结VQ-GAN离散化，与本体、动作和目标图像统一输入自回归Transformer，以行为克隆跨平台训练。新任务先少量示范微调，再由专门策略采集更多数据并并入下一代通用模型；成功检测器和可互相重置任务的…", "abstractZh": "把图像经冻结VQ-GAN离散化，与本体、动作和目标图像统一输入自回归Transformer，以行为克隆跨平台训练。新任务先少量示范微调，再由专门策略采集更多数据并并入下一代通用模型；成功检测器和可互相重置任务的策略池支持采集。\n覆盖仿真Sawyer/Panda及真实Sawyer/Panda，KUKA带三指手为训练外形态适配。通用模型评测141种训练任务变体，新任务使用100–1000示范；一般每任务至少100次，以每5000训练步25次初筛，再给选中检查点100次最终评测。", "category": "通用机器人策略 / 自我改进", "tags": [ "RoboCat", "通用机器人策略", "模仿学习", "操作与抓取" ], "directions": [ "通用机器人策略", "操作与抓取", "模仿学习" ], "tier": "classic", "experimentType": "both", "experimentNote": "覆盖仿真Sawyer/Panda及真实Sawyer/Panda，KUKA带三指手为训练外形态适配。通用模型评测141种训练任务变体，新任务使用100–1000示范；一般每任务至少100次，以每5000训练步25次初筛，再给选中检查点100次最终评测。", "robots": [ "Sawyer", "Franka Panda", "KUKA（三指手平台）" ], "robotFilters": [ "Sawyer", "Franka Panda", "KUKA" ], "robotNote": "仅列原文明确给出的实验机器人型号。", "codeStatus": "unknown", "status": "官方代码未核实", "trainingStatus": "官方代码未核实", "trainingNote": "论文和实验细节公开；MoMa 环境工具不等于 RoboCat 模型训练代码，官方完整代码状态未知。", "codeStatusNote": "本轮未核实可复现该论文的官方训练仓库；不据此断言从未发布。", "license": "未核实", "contribution": "把图像经冻结VQ-GAN离散化，与本体、动作和目标图像统一输入自回归Transformer，以行为克隆跨平台训练。新任务先少量示范微调，再由专门策略采集更多数据并并入下一代通用模型；成功检测器和可互相重置任务的策略池支持采集。", "whyUseful": "需要完整多任务数据配比、目标重标注、图像词典、专家与自主轨迹来源及检查点选择协议。应分清见过任务、微调新任务和未知形态迁移；未声称公开权重即可完整复现。", "limitations": "自改进仍需人类新任务示范、成功标签及可重置环境，不是无限制自学。实验室背景较相似，目标为图像而非语言；成功检测约90%准确，最终成绩由人工计数，不能用自动检测可靠性代替任务成绩。", "caveats": "泛化表现随任务轴变化；需要大量训练经验；本轮未核实完整官方训练代码与权重。", "evidence": [ { "url": "https://arxiv.org/abs/2306.11706", "note": "首发日期；当前标题为 Generalist，v1 标题曾用 Foundation。" }, { "url": "https://arxiv.org/html/2306.11706v1", "note": "§3、表3和附录C：仿真/实机、Sawyer、Panda、KUKA LBR IIWA14 及自举流程。" }, { "url": "https://arxiv.org/pdf/2306.11706", "note": "§2; Figure 2：图像目标、策略池和KUKA仅微调时引入。" }, { "url": "https://arxiv.org/pdf/2306.11706", "note": "§4.3; Table 2; §7：至少100回合、80/79/78%堆叠对比、受控实验室边界。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2023：展示多任务、多本体通用策略通过采集自身经验继续改进。", "freshness": "经典工作", "original": { "id": "arxiv-2306.11706", "originalSourceUrl": "https://arxiv.org/pdf/2306.11706v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2306.11706v2", "finalDownloadUrl": "https://arxiv.org/pdf/2306.11706", "sha256": "ec66c1579baaada787820abf4ce99a9a59143b583ddcc6dfc4814623d1969831", "bytes": 26404959, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:09.321507+00:00", "publicationDate": "2023/06/20", "publicationDateSourceUrl": "https://arxiv.org/abs/2306.11706", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2306.11706", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2306.11706", "pages": 60, "sourceIdentityStatus": "title_match", "sourceTitle": "RoboCat: A Self-Improving Generalist Agent for Robotic Manipulation", "sourceVersion": "2306.11706v2", "sourceVersionDate": "2023/12/22", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.094056+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2306.11706" }, "originalAnalysis": { "id": "arxiv-2306.11706", "sourceUrl": "https://arxiv.org/pdf/2306.11706", "sectionsRead": [ "§2.1–2.2", "§4.2–4.4", "§5.1–5.3", "§7" ], "methodsZh": "把图像经冻结VQ-GAN离散化，与本体、动作和目标图像统一输入自回归Transformer，以行为克隆跨平台训练。新任务先少量示范微调，再由专门策略采集更多数据并并入下一代通用模型；成功检测器和可互相重置任务的策略池支持采集。", "experimentsZh": "覆盖仿真Sawyer/Panda及真实Sawyer/Panda，KUKA带三指手为训练外形态适配。通用模型评测141种训练任务变体，新任务使用100–1000示范；一般每任务至少100次，以每5000训练步25次初筛，再给选中检查点100次最终评测。", "resultsZh": "作者报告单模型能执行跨形态堆叠、搬运和插装，并在多数真实训练任务上优于单任务视觉模型。RGB堆叠平均80%，BC-IMP79%、Gato78%，显示多任务覆盖并非每项都大幅领先；迭代数据扩大改善新任务适配。", "limitationsZh": "自改进仍需人类新任务示范、成功标签及可重置环境，不是无限制自学。实验室背景较相似，目标为图像而非语言；成功检测约90%准确，最终成绩由人工计数，不能用自动检测可靠性代替任务成绩。", "reproductionZh": "需要完整多任务数据配比、目标重标注、图像词典、专家与自主轨迹来源及检查点选择协议。应分清见过任务、微调新任务和未知形态迁移；未声称公开权重即可完整复现。", "experimentType": "both", "robots": [ "Sawyer", "Franka Panda", "KUKA（三指手平台）" ], "evidenceNotes": [ { "section/page": "§2; Figure 2", "note": "图像目标、策略池和KUKA仅微调时引入。" }, { "section/page": "§4.3; Table 2; §7", "note": "至少100回合、80/79/78%堆叠对比、受控实验室边界。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2306.03310", "title": "LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning", "titleZh": "LIBERO：终身机器人学习中的知识迁移基准", "shortTitle": "LIBERO", "date": "2023-06-05", "year": 2023, "url": "https://arxiv.org/abs/2306.03310", "codeUrl": "https://github.com/Lifelong-Robot-Learning/LIBERO", "category": "数据集与基准", "tags": [ "终身学习", "知识迁移", "语言操作", "仿真" ], "tier": "foundation", "summary": "LIBERO以Ego4D语言行为模板、初始物体分布和PDDL式目标谓词生成操作任务，专门拆分空间、物体、目标和综合知识迁移。三种策略架构配顺序学习、经验回放、EWC、PackNet和多任务训练，研究持续学新技能…", "abstractZh": "LIBERO以Ego4D语言行为模板、初始物体分布和PDDL式目标谓词生成操作任务，专门拆分空间、物体、目标和综合知识迁移。三种策略架构配顺序学习、经验回放、EWC、PackNet和多任务训练，研究持续学新技能时的遗忘与前向迁移。\nRobosuite仿真共130任务，每任务50条人工遥操作示范；前三套各10任务，LIBERO-100含90短任务预训练及10长任务。主要指标为学习速度FWT、负向后向迁移NBT及综合AUC，三种子，并改变任务顺序及语言嵌入。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "sim", "robots": [ "Franka Panda（仿真，原任务平台）" ], "codeStatus": "open", "status": "MIT 基准与训练代码公开；数据为 CC BY 4.0。", "trainingNote": "130 项任务来自仿真；人类遥操作示范不等于真实机器人采集。后续 VLA 论文常使用其部分套件。", "whyUseful": "按任务顺序保存逐任务成功矩阵与回放预算，同时报FWT/NBT/AUC；不应只报最终平均成功。明确LIBERO-90预训练是否使用，固定演示划分和三种子，语言消融应增加语义可组合性测试。", "contribution": "LIBERO以Ego4D语言行为模板、初始物体分布和PDDL式目标谓词生成操作任务，专门拆分空间、物体、目标和综合知识迁移。三种策略架构配顺序学习、经验回放、EWC、PackNet和多任务训练，研究持续学新技能时的遗忘与前向迁移。", "limitations": "原论文是持续学习基准，不是后来常见四套静态多任务成功率榜单；只在仿真，真实迁移及自然语言组合理解未得到证明。指标取保存检查点中的最佳表现，需与固定最后检查点评价区分。", "caveats": "LIBERO-100 包含 LIBERO-90 与 LIBERO-10；不能把全部 130 任务等同于常见 40 任务评测。", "evidence": [ { "url": "https://arxiv.org/abs/2306.03310", "note": "首发日期与最新版题名、摘要。" }, { "url": "https://arxiv.org/html/2306.03310v1", "note": "程序化仿真任务与实验设计。" }, { "url": "https://github.com/Lifelong-Robot-Learning/LIBERO", "note": "代码和数据许可证、任务套件。" }, { "url": "https://github.com/Lifelong-Robot-Learning/LIBERO/blob/master/libero/libero/envs/env_wrapper.py", "note": "官方环境默认 robots=[Panda]。" }, { "url": "https://arxiv.org/pdf/2306.03310", "note": "4.1–4.4; 5.1：130任务、50示范及持续学习指标定义。" }, { "url": "https://arxiv.org/pdf/2306.03310", "note": "5.2 Tables 1–3; 7：算法/架构/语言结论及仿真范围。" }, { "url": "https://arxiv.org/pdf/2306.03310v2", "note": "4.1–4.4; 5.1：130任务、50示范及持续学习指标定义。" }, { "url": "https://arxiv.org/pdf/2306.03310v2", "note": "5.2 Tables 1–3; 7：算法/架构/语言结论及仿真范围。" } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "操作与抓取", "数据集与基准" ], "robotFilters": [], "verificationNote": "已核论文、官方仿真环境代码与许可证；无实机评测。", "original": { "id": "arxiv-2306.03310", "originalSourceUrl": "https://arxiv.org/pdf/2306.03310v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2306.03310v2", "finalDownloadUrl": "https://arxiv.org/pdf/2306.03310", "sha256": "ff7d943e2eb37760df684f3ae2931a5a35ef8ea465bd387e62afe49507b8e414", "bytes": 37420885, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:17.540735+00:00", "publicationDate": "2023/06/05", "publicationDateSourceUrl": "https://arxiv.org/abs/2306.03310", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2306.03310", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2306.03310", "pages": 44, "sourceIdentityStatus": "title_match", "sourceTitle": "LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning", "sourceVersion": "2306.03310v2", "sourceVersionDate": "2023/10/14", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.129902+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2306.03310" }, "originalAnalysis": { "id": "arxiv-2306.03310", "sourceUrl": "https://arxiv.org/pdf/2306.03310v2", "sectionsRead": [ "3", "4.1–4.4", "5.1–5.2", "7", "Tables 1–3" ], "methodsZh": "LIBERO以Ego4D语言行为模板、初始物体分布和PDDL式目标谓词生成操作任务，专门拆分空间、物体、目标和综合知识迁移。三种策略架构配顺序学习、经验回放、EWC、PackNet和多任务训练，研究持续学新技能时的遗忘与前向迁移。", "experimentsZh": "Robosuite仿真共130任务，每任务50条人工遥操作示范；前三套各10任务，LIBERO-100含90短任务预训练及10长任务。主要指标为学习速度FWT、负向后向迁移NBT及综合AUC，三种子，并改变任务顺序及语言嵌入。", "resultsZh": "普通顺序学习前向迁移反而最好，所测持续学习算法都牺牲新任务学习；PackNet擅长抑制遗忘但长任务容量受限，经验回放较稳健，EWC可比朴素顺序学习更差。BERT/CLIP/GPT-2与任务ID未表现显著区别，提醒语言可能仅作为标签。", "limitationsZh": "原论文是持续学习基准，不是后来常见四套静态多任务成功率榜单；只在仿真，真实迁移及自然语言组合理解未得到证明。指标取保存检查点中的最佳表现，需与固定最后检查点评价区分。", "reproductionZh": "按任务顺序保存逐任务成功矩阵与回放预算，同时报FWT/NBT/AUC；不应只报最终平均成功。明确LIBERO-90预训练是否使用，固定演示划分和三种子，语言消融应增加语义可组合性测试。", "experimentType": "sim", "robots": [], "corrections": [ "原始基准总量130任务；常见40任务评测不代表全部原始任务。", "原文正文未明确机器人商业型号，本轮不依据基准常见实现推定型号。" ], "evidenceNotes": [ { "note": "130任务、50示范及持续学习指标定义。", "section": "4.1–4.4; 5.1" }, { "note": "算法/架构/语言结论及仿真范围。", "section": "5.2 Tables 1–3; 7" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2306.03310v2", "originalSha256": "ff7d943e2eb37760df684f3ae2931a5a35ef8ea465bd387e62afe49507b8e414" }, "experimentNote": "Robosuite仿真共130任务，每任务50条人工遥操作示范；前三套各10任务，LIBERO-100含90短任务预训练及10长任务。主要指标为学习速度FWT、负向后向迁移NBT及综合AUC，三种子，并改变任务顺序及语言嵌入。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2305.12127", "title": "DexPBT: Scaling up Dexterous Manipulation for Hand-Arm Systems with Population Based Training", "titleZh": "DexPBT：以种群训练扩展灵巧手臂系统的操作学习", "shortTitle": "DexPBT", "date": "2023-05-20", "datePrecision": "day", "year": 2023, "url": "https://arxiv.org/abs/2305.12127", "paperUrl": "https://arxiv.org/abs/2305.12127", "codeUrl": "https://github.com/isaac-sim/IsaacGymEnvs", "summary": "DexPBT以循环PPO控制7自由度Kuka臂和16自由度Allegro手，并扩展为46维双臂手系统。阶段互斥的接近、抬升、目标进度奖励避免停在目标附近刷分；以四个物体关键点统一位置与姿态误差，容差从7.5厘米…", "abstractZh": "DexPBT以循环PPO控制7自由度Kuka臂和16自由度Allegro手，并扩展为46维双臂手系统。阶段互斥的接近、抬升、目标进度奖励避免停在目标附近刷分；以四个物体关键点统一位置与姿态误差，容差从7.5厘米逐步降至1厘米。外层种群训练复制优良权重并扰动奖励和PPO超参。\n全部实验在Isaac Gym，包含重抓、抛掷、单/双臂重定向；物体为3至30厘米长方体。每V100并行8192环境，八个PBT成员与八个独立PPO种子使用相同算力，比较最佳个体；另测16/32成员规模，不是八次独立PBT重复实验。", "category": "灵巧手 / 大规模强化学习", "tags": [ "种群训练", "双臂操作", "探索", "GPU仿真" ], "directions": [ "灵巧手", "操作与抓取", "强化学习" ], "tier": "classic", "experimentType": "sim", "robots": [ "Kuka机械臂（仿真）", "Allegro Hand（仿真）" ], "robotNote": "Allegro为仿真手；机械臂原文写7-DoF KUKA，本次未进一步核实精确子型号，未推填。", "codeStatus": "open", "status": "官方PBT与操作环境代码公开；IsaacGymEnvs现为归档仓库。", "trainingNote": "官方docs/pbt.md提供去中心化种群算法、AllegroKuka任务与多GPU启动方法；需旧版Isaac Gym环境。", "contribution": "DexPBT以循环PPO控制7自由度Kuka臂和16自由度Allegro手，并扩展为46维双臂手系统。阶段互斥的接近、抬升、目标进度奖励避免停在目标附近刷分；以四个物体关键点统一位置与姿态误差，容差从7.5厘米逐步降至1厘米。外层种群训练复制优良权重并扰动奖励和PPO超参。", "whyUseful": "复现重点是rl_games、Isaac Gym、共享目录式异步PBT、变异后适应期和统一元目标。50亿步单臂约30小时、双臂约40小时是每成员时间，种群总GPU成本应单列，不能只复用最后一组超参代替完整进化过程。", "limitations": "策略直接读取物体姿态、速度和尺寸等仿真状态，真实感知尚待解决。论文没有实机验证；作者警告接近硬件极限的激烈控制可能损坏设备，动力学随机化和安全运动先验仍属未来方向。", "evidence": [ { "url": "https://arxiv.org/abs/2305.12127", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/pdf/2305.12127", "note": "正文明确16-DoF Allegro手加7-DoF KUKA臂及46-DoF双臂仿真。" }, { "url": "https://sites.google.com/view/dexpbt", "note": "作者项目页链接官方实现与PBT文档。" }, { "url": "https://github.com/isaac-sim/IsaacGymEnvs/blob/main/docs/pbt.md", "note": "官方算法与训练命令；仓库提示2026-04-14归档。" }, { "url": "https://arxiv.org/pdf/2305.12127v1", "note": "III-A/B：Kuka 7DoF + Allegro 16DoF; dual 46DoF; parallelepipeds 3–30cm; continuous-success metric cap 50." }, { "url": "https://arxiv.org/pdf/2305.12127v1", "note": "III-D 算法1/表III：Top/middle/bottom 30/40/30%; weight transfer and hyperparameter mutation; 50M-step adaptation." }, { "url": "https://arxiv.org/pdf/2305.12127v1", "note": "IV 图4–6：Equal compute best-of-eight comparison; 32-agent run total 0.32 trillion steps; >42 consecutive successes." }, { "url": "https://arxiv.org/pdf/2305.12127v1", "note": "V：Simulation-only; aggressive behavior may damage equipment; sim-to-real remains future work." } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2023：把灵巧控制的扩展重点推向种群探索与高维双臂训练系统。", "projectUrl": null, "robotFilters": [ "Allegro Hand" ], "original": { "id": "arxiv-2305.12127", "originalSourceUrl": "https://arxiv.org/pdf/2305.12127v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2305.12127v1", "finalDownloadUrl": "https://arxiv.org/pdf/2305.12127", "sha256": "66bc06ade6905f17a4dbb4b15a94cd2eb88129c56be3eb1d2fe1351d72304722", "bytes": 3619977, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:20.886921+00:00", "publicationDate": "2023/05/20", "publicationDateSourceUrl": "https://arxiv.org/abs/2305.12127", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2305.12127", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2305.12127", "pages": 11, "sourceIdentityStatus": "title_match", "sourceTitle": "DexPBT: Scaling up Dexterous Manipulation for Hand-Arm Systems with Population Based Training", "sourceVersion": "2305.12127v1", "sourceVersionDate": "2023/05/20", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.135260+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2305.12127" }, "originalAnalysis": { "id": "arxiv-2305.12127", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739393+00:00", "sourceUrl": "https://arxiv.org/pdf/2305.12127v1", "sectionsRead": [ "III-A/B 平台、物体与任务", "III-C/D PPO、奖励与PBT算法", "IV 计算预算及图4–6结果", "V 实机限制" ], "methodsZh": "DexPBT以循环PPO控制7自由度Kuka臂和16自由度Allegro手，并扩展为46维双臂手系统。阶段互斥的接近、抬升、目标进度奖励避免停在目标附近刷分；以四个物体关键点统一位置与姿态误差，容差从7.5厘米逐步降至1厘米。外层种群训练复制优良权重并扰动奖励和PPO超参。", "experimentsZh": "全部实验在Isaac Gym，包含重抓、抛掷、单/双臂重定向；物体为3至30厘米长方体。每V100并行8192环境，八个PBT成员与八个独立PPO种子使用相同算力，比较最佳个体；另测16/32成员规模，不是八次独立PBT重复实验。", "resultsZh": "作者报告普通PPO八个种子均未达到单臂重定向最终容差，PBT可达；双臂平均连续成功次数接近40，上限50。32成员、每成员100亿步时最佳个体超过42次，但总经验达3200亿步，不能将其解释为低样本方法。", "limitationsZh": "策略直接读取物体姿态、速度和尺寸等仿真状态，真实感知尚待解决。论文没有实机验证；作者警告接近硬件极限的激烈控制可能损坏设备，动力学随机化和安全运动先验仍属未来方向。", "reproductionZh": "复现重点是rl_games、Isaac Gym、共享目录式异步PBT、变异后适应期和统一元目标。50亿步单臂约30小时、双臂约40小时是每成员时间，种群总GPU成本应单列，不能只复用最后一组超参代替完整进化过程。", "evidenceNotes": [ { "section/page": "III-A/B", "note": "Kuka 7DoF + Allegro 16DoF; dual 46DoF; parallelepipeds 3–30cm; continuous-success metric cap 50." }, { "section/page": "III-D 算法1/表III", "note": "Top/middle/bottom 30/40/30%; weight transfer and hyperparameter mutation; 50M-step adaptation." }, { "section/page": "IV 图4–6", "note": "Equal compute best-of-eight comparison; 32-agent run total 0.32 trillion steps; >42 consecutive successes." }, { "section/page": "V", "note": "Simulation-only; aggressive behavior may damage equipment; sim-to-real remains future work." } ], "experimentType": "sim", "robots": [ "Kuka机械臂（仿真）", "Allegro Hand（仿真）" ], "corrections": [ "补充原文Kuka机械臂；明确不是实机手臂实验。" ], "sourceVersion": "2305.12127v1", "originalSha256": "66bc06ade6905f17a4dbb4b15a94cd2eb88129c56be3eb1d2fe1351d72304722" }, "experimentNote": "全部实验在Isaac Gym，包含重抓、抛掷、单/双臂重定向；物体为3至30厘米长方体。每V100并行8192环境，八个PBT成员与八个独立PPO种子使用相同算力，比较最佳个体；另测16/32成员规模，不是八次独立PBT重复实验。", "analysisVerifiedAt": "2026-10-04T13:47:58.739393+00:00" }, { "id": "arxiv-2305.05706", "title": "DexArt: Benchmarking Generalizable Dexterous Manipulation with Articulated Objects", "titleZh": "DexArt：可泛化关节物体灵巧操作基准", "shortTitle": "DexArt", "date": "2023-05-09", "datePrecision": "day", "year": 2023, "url": "https://arxiv.org/abs/2305.05706", "paperUrl": "https://arxiv.org/abs/2305.05706", "codeUrl": "https://github.com/Kami-code/dexart-release", "summary": "DexArt构建多指手操作关节物体基准，以部分点云、本体状态和根据机器人模型生成的补充点云输入PointNet与PPO。比较分类、部件分割、重建和SimSiam预训练；奖励分为接近、建立接触和推进物体关节阶段。", "abstractZh": "DexArt构建多指手操作关节物体基准，以部分点云、本体状态和根据机器人模型生成的补充点云输入PointNet与PPO。比较分类、部件分割、重建和SimSiam预训练；奖励分为接近、建立接触和推进物体关节阶段。\nSAPIEN中使用xArm6与16自由度Allegro Hand，四任务水龙头、桶、笔记本、马桶，共82物体，按类别划分已见/未见实例。每配置三随机种子，评估成功、回报、训练物体数量、PointNet容量和相机视角扰动。", "category": "灵巧手 / 数据集与基准", "tags": [ "关节物体", "点云", "视觉预训练", "泛化基准" ], "directions": [ "灵巧手", "操作与抓取", "数据集与基准" ], "tier": "classic", "experimentType": "sim", "robots": [ "xArm6", "Allegro Hand" ], "robotNote": "xArm6和Allegro均为仿真模型；不列为实机验证。", "codeStatus": "open", "status": "官方环境、PPO训练、视觉预训练和评估代码公开。", "trainingNote": "下载对象与预训练资产后运行对应四任务；官方脚本区分训练实例和未见测试实例。", "contribution": "DexArt构建多指手操作关节物体基准，以部分点云、本体状态和根据机器人模型生成的补充点云输入PointNet与PPO。比较分类、部件分割、重建和SimSiam预训练；奖励分为接近、建立接触和推进物体关节阶段。", "whyUseful": "复现需保持82物体身份划分、点云裁剪采样、想象机器人点云和接触奖励，避免用未见对象进行部件监督造成泄漏。对比预训练方法应分清DAM与更广PMM数据，曲线的标准差/标准误口径亦不同。", "limitations": "全部为仿真，xArm6/Allegro是仿真本体，输入可在现实获得不等于已完成实机部署。资产经过人工选择、比例和初态标注，任务奖励有工程先验；真实深度噪声、接触与硬件安全未被这组结果验证。", "evidence": [ { "url": "https://arxiv.org/abs/2305.05706", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/pdf/2305.05706", "note": "§3.2明确SAPIEN、XArm6和Allegro模型；四任务均为仿真。" }, { "url": "https://www.chenbao.tech/dexart/", "note": "作者页概述四任务、PointNet预训练与泛化研究。" }, { "url": "https://github.com/Kami-code/dexart-release", "note": "官方训练、测试集评估和视觉预训练命令。" }, { "url": "https://arxiv.org/pdf/2305.05706", "note": "§3.2 / Table 1：SAPIEN、xArm6+Allegro、四类82资产；无实体实验。" }, { "url": "https://arxiv.org/pdf/2305.05706", "note": "Table 2：DAM部件分割未见成功58/76/60/55%。" }, { "url": "https://arxiv.org/pdf/2305.05706", "note": "Fig.6 caption：成功曲线为三种子标准差，回报曲线用标准误。" }, { "url": "https://arxiv.org/pdf/2305.05706", "note": "PDF p.6, Table 2 (visual inspection)：已逐列核对seen/unseen及均值±标准差表头，确认58/76/60/55为未见实例列。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2023：灵巧操作评估从单物体成功率转向类别级关节物体泛化。", "projectUrl": null, "robotFilters": [], "original": { "id": "arxiv-2305.05706", "originalSourceUrl": "https://arxiv.org/pdf/2305.05706v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2305.05706v1", "finalDownloadUrl": "https://arxiv.org/pdf/2305.05706", "sha256": "44354c0496a14f9ef90e19cd93f1521075adb6daff5f0a882cc045a649fbff7e", "bytes": 4578047, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:21.152049+00:00", "publicationDate": "2023/05/09", "publicationDateSourceUrl": "https://arxiv.org/abs/2305.05706", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2305.05706", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2305.05706", "pages": 11, "sourceIdentityStatus": "title_match", "sourceTitle": "DexArt: Benchmarking Generalizable Dexterous Manipulation with Articulated Objects", "sourceVersion": "2305.05706v1", "sourceVersionDate": "2023/05/09", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.139975+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2305.05706" }, "originalAnalysis": { "id": "arxiv-2305.05706", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090138+00:00", "sourceUrl": "https://arxiv.org/pdf/2305.05706", "sectionsRead": [ "§3", "§4", "§5.1–5.3", "Tables 1–3", "Figs.6–9" ], "methodsZh": "DexArt构建多指手操作关节物体基准，以部分点云、本体状态和根据机器人模型生成的补充点云输入PointNet与PPO。比较分类、部件分割、重建和SimSiam预训练；奖励分为接近、建立接触和推进物体关节阶段。", "experimentsZh": "SAPIEN中使用xArm6与16自由度Allegro Hand，四任务水龙头、桶、笔记本、马桶，共82物体，按类别划分已见/未见实例。每配置三随机种子，评估成功、回报、训练物体数量、PointNet容量和相机视角扰动。", "resultsZh": "作者表2部件分割DAM预训练在未见四任务为58%、76%、60%、55%；无预训练28%、56%、41%、46%。结果支持任务相关功能部件表示，但普通类别预训练未普遍带来收益；较小PointNet还可能更易训练和泛化。", "limitationsZh": "全部为仿真，xArm6/Allegro是仿真本体，输入可在现实获得不等于已完成实机部署。资产经过人工选择、比例和初态标注，任务奖励有工程先验；真实深度噪声、接触与硬件安全未被这组结果验证。", "reproductionZh": "复现需保持82物体身份划分、点云裁剪采样、想象机器人点云和接触奖励，避免用未见对象进行部件监督造成泄漏。对比预训练方法应分清DAM与更广PMM数据，曲线的标准差/标准误口径亦不同。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "§3.2 / Table 1", "note": "SAPIEN、xArm6+Allegro、四类82资产；无实体实验。" }, { "section/page": "Table 2", "note": "DAM部件分割未见成功58/76/60/55%。" }, { "section/page": "Fig.6 caption", "note": "成功曲线为三种子标准差，回报曲线用标准误。" }, { "section/page": "PDF p.6, Table 2 (visual inspection)", "note": "已逐列核对seen/unseen及均值±标准差表头，确认58/76/60/55为未见实例列。" } ], "corrections": { "robotNote": "xArm6与Allegro Hand仅作为SAPIEN仿真配置。" } }, "experimentNote": "SAPIEN中使用xArm6与16自由度Allegro Hand，四任务水龙头、桶、笔记本、马桶，共82物体，按类别划分已见/未见实例。每配置三随机种子，评估成功、回报、训练物体数量、PointNet容量和相机视角扰动。", "analysisVerifiedAt": "2026-10-04T13:50:35.090138+00:00" }, { "id": "arxiv-2304.13705", "title": "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware", "titleZh": "用低成本硬件学习精细双臂操作", "shortTitle": "ACT / ALOHA", "date": "2023-04-23", "datePrecision": "day", "year": 2023, "url": "https://arxiv.org/abs/2304.13705", "codeUrl": "https://github.com/tonyzhaozh/act", "category": "双臂操作", "tags": [ "动作分块", "遥操作", "模仿学习" ], "tier": "classic", "summary": "ALOHA以小型WidowX领导臂直接映射关节到ViperX双跟随臂，配四路相机和改装夹爪。ACT使用条件VAE Transformer一次预测动作块，执行时每步重新查询，并对重叠预测做时间集成，缓解误差累积与…", "abstractZh": "ALOHA以小型WidowX领导臂直接映射关节到ViperX双跟随臂，配四路相机和改装夹爪。ACT使用条件VAE Transformer一次预测动作块，执行时每步重新查询，并对重叠预测做时间集成，缓解误差累积与示范中暂停造成的非马尔可夫性。\n设计两项MuJoCo任务及六项真实精细双臂操作；仿真区分脚本/人工示范，三种子每次50回合，真实主对比一训练种子、25次试验。比较BC、BeT、VINN等，并消融动作块长度、时间集成和VAE。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "ALOHA", "Trossen ViperX", "Trossen WidowX" ], "codeStatus": "open", "status": "官方ACT与ALOHA代码公开；硬件教程可获取。", "trainingNote": "离线模仿学习；ViperX为执行臂，WidowX为遥操作主臂。", "whyUseful": "优先恢复领导—跟随标定、相机同步、关节目标而非增量动作、动作块及集成权重；按阶段和完整成功分别计分。训练需任务示范，本次未装配或执行代码。", "contribution": "ALOHA以小型WidowX领导臂直接映射关节到ViperX双跟随臂，配四路相机和改装夹爪。ACT使用条件VAE Transformer一次预测动作块，执行时每步重新查询，并对重叠预测做时间集成，缓解误差累积与示范中暂停造成的非马尔可夫性。", "limitations": "策略逐任务训练，评测初始状态和工作区有限；低成本臂原始绝对精度不高，成功依靠视觉反馈和任务设置。单实机训练种子难估计训练方差，长块减少反应性，不能等同通用语言策略。", "caveats": "论文中的分钟数是示范数据时长，不等于采集总耗时或训练耗时。", "evidence": [ { "url": "https://arxiv.org/abs/2304.13705", "note": "首发日期与ACT摘要。" }, { "url": "https://arxiv.org/pdf/2304.13705", "note": "第III节及参考文献1/2：ViperX 300、WidowX 250；第V节仿真与真机。" }, { "url": "https://tonyzhaozh.github.io/aloha/", "note": "作者项目页链接ACT和ALOHA。" }, { "url": "https://github.com/tonyzhaozh/act", "note": "MIT许可；训练评测和两个仿真环境。" }, { "url": "https://github.com/tonyzhaozh/aloha", "note": "遥操作与真实机器人代码。" }, { "url": "https://arxiv.org/pdf/2304.13705", "note": "§III–IV; Algorithm 2：低成本双臂关节映射、CVAE及时间集成。" }, { "url": "https://arxiv.org/pdf/2304.13705", "note": "§V–VI Tables I–II：25次实机主比较、88%/96%及动作块消融。" } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "操作与抓取", "模仿学习" ], "robotFilters": [ "ALOHA", "Trossen ViperX", "WidowX" ], "verificationNote": "2026-10-04核对原论文、硬件角色与官方代码；不混入Mobile ALOHA。", "original": { "id": "arxiv-2304.13705", "originalSourceUrl": "https://arxiv.org/pdf/2304.13705v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2304.13705v1", "finalDownloadUrl": "https://arxiv.org/pdf/2304.13705", "sha256": "2e2fe25860f5f9cee9e655a2714e9f2264a7a5078bcd267c16d3f39af461345a", "bytes": 5353485, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:30.513035+00:00", "publicationDate": "2023/04/23", "publicationDateSourceUrl": "https://arxiv.org/abs/2304.13705", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2304.13705", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2304.13705", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware", "sourceVersion": "2304.13705v1", "sourceVersionDate": "2023/04/23", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.144295+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2304.13705" }, "originalAnalysis": { "id": "arxiv-2304.13705", "sourceUrl": "https://arxiv.org/pdf/2304.13705", "sectionsRead": [ "§III–IV", "§V–VI", "Tables I–II", "Algorithms 1–2" ], "methodsZh": "ALOHA以小型WidowX领导臂直接映射关节到ViperX双跟随臂，配四路相机和改装夹爪。ACT使用条件VAE Transformer一次预测动作块，执行时每步重新查询，并对重叠预测做时间集成，缓解误差累积与示范中暂停造成的非马尔可夫性。", "experimentsZh": "设计两项MuJoCo任务及六项真实精细双臂操作；仿真区分脚本/人工示范，三种子每次50回合，真实主对比一训练种子、25次试验。比较BC、BeT、VINN等，并消融动作块长度、时间集成和VAE。", "resultsZh": "作者报告开拉链袋和插电池成功88%、96%，对照难以完成后段。禁用集成时，仿真四设置平均成功从单步约1%升到100步块约44%，块过长又下降。正文实际重建使用L1，算法伪码写MSE，复现须以实现/正文配置核对。", "limitationsZh": "策略逐任务训练，评测初始状态和工作区有限；低成本臂原始绝对精度不高，成功依靠视觉反馈和任务设置。单实机训练种子难估计训练方差，长块减少反应性，不能等同通用语言策略。", "reproductionZh": "优先恢复领导—跟随标定、相机同步、关节目标而非增量动作、动作块及集成权重；按阶段和完整成功分别计分。训练需任务示范，本次未装配或执行代码。", "experimentType": "both", "robots": [ "ALOHA", "Trossen ViperX", "Trossen WidowX" ], "evidenceNotes": [ { "section/page": "§III–IV; Algorithm 2", "note": "低成本双臂关节映射、CVAE及时间集成。" }, { "section/page": "§V–VI Tables I–II", "note": "25次实机主比较、88%/96%及动作块消融。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "设计两项MuJoCo任务及六项真实精细双臂操作；仿真区分脚本/人工示范，三种子每次50回合，真实主对比一训练种子、25次试验。比较BC、BeT、VINN等，并消融动作块长度、时间集成和VAE。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2303.04137", "title": "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion", "titleZh": "扩散策略：通过动作扩散学习视觉运动策略", "shortTitle": "Diffusion Policy", "date": "2023-03-07", "datePrecision": "day", "year": 2023, "url": "https://arxiv.org/abs/2303.04137v4", "codeUrl": "https://github.com/real-stanford/diffusion_policy", "category": "模仿学习", "tags": [ "动作扩散", "动作序列", "视觉控制" ], "tier": "classic", "summary": "Diffusion Policy将视觉条件动作分布表示为逐步去噪的动作序列，使用滚动时域执行部分动作后重新观测。视觉特征只作条件而不与动作共同去噪；比较时序卷积U-Net和Transformer、位置/速度动作…", "abstractZh": "Diffusion Policy将视觉条件动作分布表示为逐步去噪的动作序列，使用滚动时域执行部分动作后重新观测。视觉特征只作条件而不与动作共同去噪；比较时序卷积U-Net和Transformer、位置/速度动作与视觉预训练，避免确定性回归平均多个行为模式。\n四个模拟基准的12任务覆盖Push-T、Robomimic和Franka Kitchen等；UR5真实Push-T，以及Panda的翻转、酱料倒取/涂抹等动态或多阶段任务。实机Push-T保留停顿动作，以末态IoU和移至结束区共同检验细调及阶段切换。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "Universal Robots UR5", "Franka Emika Panda" ], "codeStatus": "open", "status": "官方仿真与真实机器人代码、数据及训练配置公开。", "trainingNote": "行为克隆式去噪训练；推理成本随采样步骤变化。", "whyUseful": "从公开基准固定观察/动作窗口、执行段长度、归一化和DDIM步数，先复现位置控制与序列预测消融。真实10Hz指令插值到125Hz，不应把插值频率算成策略推理频率；同时测延迟和闭环成功。", "contribution": "Diffusion Policy将视觉条件动作分布表示为逐步去噪的动作序列，使用滚动时域执行部分动作后重新观测。视觉特征只作条件而不与动作共同去噪；比较时序卷积U-Net和Transformer、位置/速度动作与视觉预训练，避免确定性回归平均多个行为模式。", "limitations": "行为克隆仍受示范覆盖和质量限制；迭代采样延迟高于简单策略，高频任务可能不足。仿真从多个检查点选择最佳、跨三种子汇报的协议应复核；真实强鲁棒扰动展示有定性成分。", "caveats": "日期为首次预印本；内容固定RSS 2023 v4，避免混入2024新增双臂实验。", "evidence": [ { "url": "https://arxiv.org/abs/2303.04137", "note": "首发日期；最新版本为2024扩展。" }, { "url": "https://arxiv.org/pdf/2303.04137v4", "note": "RSS版本图2真实UR5与Franka，正文明确Franka Panda。" }, { "url": "https://diffusion-policy.cs.columbia.edu/", "note": "区分RSS 2023与IJRR 2024，并提供代码。" }, { "url": "https://github.com/real-stanford/diffusion_policy", "note": "官方仓库当前重定向目标及训练/真实实验说明。" }, { "url": "https://arxiv.org/pdf/2303.04137", "note": "II–III; V：动作去噪、滚动执行、设计消融及仿真协议。" }, { "url": "https://arxiv.org/pdf/2303.04137", "note": "VI Table V; VIII：UR5 Push-T95%、真实平台与延迟限制。" }, { "url": "https://arxiv.org/pdf/2303.04137v4", "note": "II–III; V：动作去噪、滚动执行、设计消融及仿真协议。" }, { "url": "https://arxiv.org/pdf/2303.04137v4", "note": "VI Table V; VIII：UR5 Push-T95%、真实平台与延迟限制。" } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "模仿学习" ], "robotFilters": [ "Universal Robots UR5", "Franka Panda" ], "verificationNote": "2026-10-04核对版本、硬件与官方仓库；未由Franka Kitchen推断实体硬件。", "original": { "id": "arxiv-2303.04137", "originalSourceUrl": "https://arxiv.org/pdf/2303.04137v4", "versionedOriginalUrl": "https://arxiv.org/pdf/2303.04137v4", "finalDownloadUrl": "https://arxiv.org/pdf/2303.04137v4", "sha256": "75a57946e2cef790775ca5c667b00befc8ea7d64c3a02c5ced3e3d173354640c", "bytes": 5834180, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:30.683550+00:00", "publicationDate": "2023/03/07", "publicationDateSourceUrl": "https://arxiv.org/abs/2303.04137v4", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2303.04137v4", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2303.04137v4", "pages": 16, "sourceIdentityStatus": "title_match", "sourceTitle": "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion", "sourceVersion": "2303.04137v4", "sourceVersionDate": "2023/06/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.149809+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2303.04137v4" }, "originalAnalysis": { "id": "arxiv-2303.04137", "sourceUrl": "https://arxiv.org/pdf/2303.04137v4", "sectionsRead": [ "II–IV", "V-A–B", "VI-A–B", "VIII", "Tables I–V" ], "methodsZh": "Diffusion Policy将视觉条件动作分布表示为逐步去噪的动作序列，使用滚动时域执行部分动作后重新观测。视觉特征只作条件而不与动作共同去噪；比较时序卷积U-Net和Transformer、位置/速度动作与视觉预训练，避免确定性回归平均多个行为模式。", "experimentsZh": "四个模拟基准的12任务覆盖Push-T、Robomimic和Franka Kitchen等；UR5真实Push-T，以及Panda的翻转、酱料倒取/涂抹等动态或多阶段任务。实机Push-T保留停顿动作，以末态IoU和移至结束区共同检验细调及阶段切换。", "resultsZh": "真实Push-T端到端卷积版本95%成功、IoU0.80，人工0.84；LSTM-GMM最佳20%、IBC0%。位置控制和动作块设计共同带来收益，Transformer或预训练视觉编码并不在每项都更优，单看“用了扩散”无法解释全部差异。", "limitationsZh": "行为克隆仍受示范覆盖和质量限制；迭代采样延迟高于简单策略，高频任务可能不足。仿真从多个检查点选择最佳、跨三种子汇报的协议应复核；真实强鲁棒扰动展示有定性成分。", "reproductionZh": "从公开基准固定观察/动作窗口、执行段长度、归一化和DDIM步数，先复现位置控制与序列预测消融。真实10Hz指令插值到125Hz，不应把插值频率算成策略推理频率；同时测延迟和闭环成功。", "experimentType": "both", "robots": [ "Universal Robots UR5", "Franka Emika Panda" ], "corrections": [ "真实Push-T为10Hz策略动作插值125Hz执行；不是125Hz视觉推理。" ], "evidenceNotes": [ { "note": "动作去噪、滚动执行、设计消融及仿真协议。", "section": "II–III; V" }, { "note": "UR5 Push-T95%、真实平台与延迟限制。", "section": "VI Table V; VIII" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2303.04137v4", "originalSha256": "75a57946e2cef790775ca5c667b00befc8ea7d64c3a02c5ced3e3d173354640c" }, "experimentNote": "四个模拟基准的12任务覆盖Push-T、Robomimic和Franka Kitchen等；UR5真实Push-T，以及Panda的翻转、酱料倒取/涂抹等动态或多阶段任务。实机Push-T保留停顿动作，以末态IoU和移至结束区共同检验细调及阶段切换。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2303.03378", "title": "PaLM-E: An Embodied Multimodal Language Model", "titleZh": "PaLM-E：具身多模态语言模型", "shortTitle": "PaLM-E", "date": "2023-03-06", "year": 2023, "url": "https://arxiv.org/abs/2303.03378", "codeUrl": null, "category": "具身推理与规划", "tags": [ "多模态", "语言规划", "传感器接地", "跨任务迁移" ], "tier": "foundation", "summary": "PaLM-E把图像、连续状态或对象中心三维表示映射到语言嵌入维度，插入文本序列，以自回归语言目标共同训练。模型输出文字计划或回答；真实动作由已有语言条件低层控制器执行，再把新图像反馈给高层，不是直接生成关节动作…", "abstractZh": "PaLM-E把图像、连续状态或对象中心三维表示映射到语言嵌入维度，插入文本序列，以自回归语言目标共同训练。模型输出文字计划或回答；真实动作由已有语言条件低层控制器执行，再把新图像反馈给高层，不是直接生成关节动作的端到端控制策略。\n作者比较TAMP仿真规划、Language Table积木排序、真实桌面与厨房移动操作，并混合通用视觉问答/语言任务。消融预训练、冻结语言模型、输入表示及跨任务数据；TAMP低数据评测每类规划只用320例，桌面规划比较10至80例示范。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "SayCan移动操作平台（实机，商品型号未单列）", "Language Table桌面平台（实机与仿真，型号未单列）", "TAMP仿真操作机器人" ], "codeStatus": "unknown", "status": "已核论文与作者项目页；官方实现和权重公开性未确认。", "trainingNote": "需预训练 PaLM、视觉编码器及低层技能策略；最大模型为 562B 参数，不能据此认定所有实机实验均用该规模。", "whyUseful": "复现需PaLM/视觉预训练权重、机器人混合数据、编码器及RT-1等执行器，不能仅复现提示词。应将高层规划、可执行性判断、底层控制与最终任务成功分别计分，固定冻结策略和数据规模，避免把不同模型尺寸结果混合。", "contribution": "PaLM-E把图像、连续状态或对象中心三维表示映射到语言嵌入维度，插入文本序列，以自回归语言目标共同训练。模型输出文字计划或回答；真实动作由已有语言条件低层控制器执行，再把新图像反馈给高层，不是直接生成关节动作的端到端控制策略。", "limitations": "能力受已有低层技能词汇与执行可靠性约束，生成计划没有硬约束验证。小模型多模态微调会遗忘语言能力；跨域混合有帮助但不能推导任意新机器人零样本控制，原文未单列实机商品型号。", "caveats": "机器人名称按原文环境保留，未由图片推断商业型号。", "evidence": [ { "url": "https://arxiv.org/abs/2303.03378", "note": "首发日期、摘要与模型规模。" }, { "url": "https://arxiv.org/html/2303.03378v1", "note": "第 6 节与附录 B.2 明确真实控制回路及 Language-Table 仿真评测。" }, { "url": "https://palm-e.github.io/", "note": "两个真实机器人平台的闭环规划演示。" }, { "url": "https://arxiv.org/pdf/2303.03378v1", "note": "3：Text output conditions existing low-level policies; model replans using new observations." }, { "url": "https://arxiv.org/pdf/2303.03378v1", "note": "6.2 表1：320 examples/planning task; ViT-4B single-robot vs full-mixture results, frozen LLM." }, { "url": "https://arxiv.org/pdf/2303.03378v1", "note": "6.3–6.4：Tabletop language subgoals 1Hz, low-level 5Hz; real mobile manipulation assessed qualitatively using 2912 training sequences." }, { "url": "https://arxiv.org/pdf/2303.03378v1", "note": "表4/5：0.91 failure-detection F1, not robot task success; 562B OK-VQA 66.1." } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "视觉语言动作" ], "robotFilters": [], "verificationNote": "已核论文实验部分与官方演示；代码状态未确认。", "original": { "id": "arxiv-2303.03378", "originalSourceUrl": "https://arxiv.org/pdf/2303.03378v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2303.03378v1", "finalDownloadUrl": "https://arxiv.org/pdf/2303.03378", "sha256": "535cc3c7fe1ae7428b9ac7d6a3738a0529975aadc372b2229ea0cb2ce724576a", "bytes": 10707237, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:35.734824+00:00", "publicationDate": "2023/03/06", "publicationDateSourceUrl": "https://arxiv.org/abs/2303.03378", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2303.03378", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2303.03378", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "PaLM-E: An Embodied Multimodal Language Model", "sourceVersion": "2303.03378v1", "sourceVersionDate": "2023/03/06", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.158836+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2303.03378" }, "originalAnalysis": { "id": "arxiv-2303.03378", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739394+00:00", "sourceUrl": "https://arxiv.org/pdf/2303.03378v1", "sectionsRead": [ "3 多模态嵌入与高层控制闭环", "4 状态、ViT、OSRT编码", "6.2–6.6 TAMP、Language Table、移动操作及表1–5", "7–8 迁移与遗忘讨论" ], "methodsZh": "PaLM-E把图像、连续状态或对象中心三维表示映射到语言嵌入维度，插入文本序列，以自回归语言目标共同训练。模型输出文字计划或回答；真实动作由已有语言条件低层控制器执行，再把新图像反馈给高层，不是直接生成关节动作的端到端控制策略。", "experimentsZh": "作者比较TAMP仿真规划、Language Table积木排序、真实桌面与厨房移动操作，并混合通用视觉问答/语言任务。消融预训练、冻结语言模型、输入表示及跨任务数据；TAMP低数据评测每类规划只用320例，桌面规划比较10至80例示范。", "resultsZh": "表1中ViT全混合数据的两类TAMP成功率74.1%/74.6%，仅单机器人数据为30.6%/32.9%，支持正迁移；OSRT为82.5%/76.2%。真实厨房主要是定性长程展示；失败检测0.91为F1，不能当作实机任务成功率。562B模型OK-VQA为66.1。", "limitationsZh": "能力受已有低层技能词汇与执行可靠性约束，生成计划没有硬约束验证。小模型多模态微调会遗忘语言能力；跨域混合有帮助但不能推导任意新机器人零样本控制，原文未单列实机商品型号。", "reproductionZh": "复现需PaLM/视觉预训练权重、机器人混合数据、编码器及RT-1等执行器，不能仅复现提示词。应将高层规划、可执行性判断、底层控制与最终任务成功分别计分，固定冻结策略和数据规模，避免把不同模型尺寸结果混合。", "evidenceNotes": [ { "section/page": "3", "note": "Text output conditions existing low-level policies; model replans using new observations." }, { "section/page": "6.2 表1", "note": "320 examples/planning task; ViT-4B single-robot vs full-mixture results, frozen LLM." }, { "section/page": "6.3–6.4", "note": "Tabletop language subgoals 1Hz, low-level 5Hz; real mobile manipulation assessed qualitatively using 2912 training sequences." }, { "section/page": "表4/5", "note": "0.91 failure-detection F1, not robot task success; 562B OK-VQA 66.1." } ], "experimentType": "both", "robots": [ "SayCan移动操作平台（实机，商品型号未单列）", "Language Table桌面平台（实机与仿真，型号未单列）", "TAMP仿真操作机器人" ], "corrections": [ "区分输出文字计划的高层策略与低层动作模型。" ], "sourceVersion": "2303.03378v1", "originalSha256": "535cc3c7fe1ae7428b9ac7d6a3738a0529975aadc372b2229ea0cb2ce724576a" }, "experimentNote": "作者比较TAMP仿真规划、Language Table积木排序、真实桌面与厨房移动操作，并混合通用视觉问答/语言任务。消融预训练、冻结语言模型、输入表示及跨任务数据；TAMP低数据评测每类规划只用320例，桌面规划比较10至80例示范。", "analysisVerifiedAt": "2026-10-04T13:47:58.739394+00:00" }, { "id": "arxiv-2303.03381", "title": "Real-World Humanoid Locomotion with Reinforcement Learning", "titleZh": "用强化学习实现真实世界人形机器人运动", "shortTitle": "Humanoid Transformer", "date": "2023-03-06", "datePrecision": "day", "year": 2023, "url": "https://arxiv.org/abs/2303.03381", "paperUrl": "https://arxiv.org/abs/2303.03381", "codeUrl": null, "summary": "先在特权状态下训练教师策略，再让因果Transformer根据本体观测和既往动作历史学习学生，以教师模仿加强化学习联合优化。Isaac Gym用虚拟弹簧近似Digit闭链机构，随机化动力学、地形、噪声和延迟，部…", "abstractZh": "先在特权状态下训练教师策略，再让因果Transformer根据本体观测和既往动作历史学习学生，以教师模仿加强化学习联合优化。Isaac Gym用虚拟弹簧近似Digit闭链机构，随机化动力学、地形、噪声和延迟，部署前在Agility仿真器筛除不安全策略。\n实体Agility Digit约1.6米、45公斤，政策50Hz、关节PD1kHz；室内测试外力、杂物地面、斜坡和五种携带载荷，室外覆盖草地、路面等。另在可控仿真比较公司控制器，并消融网络、上下文长度和训练目标。", "category": "人形机器人 / 运动控制", "tags": [ "Transformer", "本体历史", "上下文适应", "Sim2Real" ], "directions": [ "人形机器人", "运动控制", "强化学习" ], "tier": "classic", "experimentType": "both", "robots": [ "Agility Robotics Digit" ], "robotNote": "", "codeStatus": "unknown", "status": "论文与作者项目公开；本次未核实作者完整训练代码。", "trainingNote": "仿真中进行大规模RL与策略学习，再以高保真仿真检查策略；需要Digit动力学和部署接口。", "contribution": "先在特权状态下训练教师策略，再让因果Transformer根据本体观测和既往动作历史学习学生，以教师模仿加强化学习联合优化。Isaac Gym用虚拟弹簧近似Digit闭链机构，随机化动力学、地形、噪声和延迟，部署前在Agility仿真器筛除不安全策略。", "whyUseful": "复现需Digit机械/驱动参数、闭链弹簧子步、随机化和教师学生流程，以及高保真仿真预筛选。所谓零样本实机指策略未实机微调，不意味着无需机器人模型和传感器知识；本轮未独立执行。", "limitations": "无外部视觉感知，依赖接触后的本体反馈；未观测跌倒没有给出统一行走距离分母，不能转成安全概率。作者指出左右不对称、速度跟踪误差以及强拉拽仍可导致跌倒，闭链仿真近似也是转移限制。", "evidence": [ { "url": "https://arxiv.org/abs/2303.03381", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/pdf/2303.03381", "note": "正文明确Digit机器人、户外部署与仿真评估流程。" }, { "url": "https://learning-humanoid-locomotion.github.io/", "note": "作者项目说明因果Transformer、训练规模与实机验证；页面未提供代码入口。" }, { "url": "https://humanoid-transformer.github.io/", "note": "原始项目题名与同一arXiv编号对应关系。" }, { "url": "https://arxiv.org/pdf/2303.03381", "note": "Outdoor deployment：作者一周室外未跌倒的观察，不是正式失效率估计。" }, { "url": "https://arxiv.org/pdf/2303.03381", "note": "Indoor/simulation comparison：不稳定木板未在真机比较。" }, { "url": "https://arxiv.org/pdf/2303.03381", "note": "Sim-to-real / Limitations：50Hz/1kHz；Agility仿真筛选不更新参数；强扰动仍会跌倒。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2023：Transformer历史建模与仿真RL结合，进入全尺寸人形机器人实机行走。", "projectUrl": null, "robotFilters": [ "Agility Robotics Digit" ], "original": { "id": "arxiv-2303.03381", "originalSourceUrl": "https://arxiv.org/pdf/2303.03381v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2303.03381v2", "finalDownloadUrl": "https://arxiv.org/pdf/2303.03381", "sha256": "9b310e368242b5d3856bcbe55990261fdd2072fba966c5f9887cc896c5e5ee5a", "bytes": 9586709, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:40.921731+00:00", "publicationDate": "2023/03/06", "publicationDateSourceUrl": "https://arxiv.org/abs/2303.03381", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2303.03381", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2303.03381", "pages": 32, "sourceIdentityStatus": "title_match", "sourceTitle": "Real-World Humanoid Locomotion with Reinforcement Learning", "sourceVersion": "2303.03381v2", "sourceVersionDate": "2023/12/14", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.166464+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2303.03381" }, "originalAnalysis": { "id": "arxiv-2303.03381", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090140+00:00", "sourceUrl": "https://arxiv.org/pdf/2303.03381", "sectionsRead": [ "Results", "Indoor and simulation experiments", "Materials and Methods", "Fig.7–8", "Limitations" ], "methodsZh": "先在特权状态下训练教师策略，再让因果Transformer根据本体观测和既往动作历史学习学生，以教师模仿加强化学习联合优化。Isaac Gym用虚拟弹簧近似Digit闭链机构，随机化动力学、地形、噪声和延迟，部署前在Agility仿真器筛除不安全策略。", "experimentsZh": "实体Agility Digit约1.6米、45公斤，政策50Hz、关节PD1kHz；室内测试外力、杂物地面、斜坡和五种携带载荷，室外覆盖草地、路面等。另在可控仿真比较公司控制器，并消融网络、上下文长度和训练目标。", "resultsZh": "作者报告一周整日室外测试未观察到跌倒，展示脚部卡住后的恢复和随环境调整步态；三类训练消融支持Transformer、较长历史和联合目标。但不稳定木板的控制器对照只做仿真，未因危险而搬到真机。", "limitationsZh": "无外部视觉感知，依赖接触后的本体反馈；未观测跌倒没有给出统一行走距离分母，不能转成安全概率。作者指出左右不对称、速度跟踪误差以及强拉拽仍可导致跌倒，闭链仿真近似也是转移限制。", "reproductionZh": "复现需Digit机械/驱动参数、闭链弹簧子步、随机化和教师学生流程，以及高保真仿真预筛选。所谓零样本实机指策略未实机微调，不意味着无需机器人模型和传感器知识；本轮未独立执行。", "experimentType": "both", "robots": [ "Agility Robotics Digit" ], "evidenceNotes": [ { "section/page": "Outdoor deployment", "note": "作者一周室外未跌倒的观察，不是正式失效率估计。" }, { "section/page": "Indoor/simulation comparison", "note": "不稳定木板未在真机比较。" }, { "section/page": "Sim-to-real / Limitations", "note": "50Hz/1kHz；Agility仿真筛选不更新参数；强扰动仍会跌倒。" } ] }, "experimentNote": "实体Agility Digit约1.6米、45公斤，政策50Hz、关节PD1kHz；室内测试外力、杂物地面、斜坡和五种携带载荷，室外覆盖草地、路面等。另在可控仿真比较公司控制器，并消融网络、上下文长度和训练目标。", "analysisVerifiedAt": "2026-10-04T13:50:35.090140+00:00" }, { "id": "arxiv-2302.12422", "title": "MimicPlay: Long-Horizon Imitation Learning by Watching Human Play", "date": "2023-02-24", "url": "https://arxiv.org/abs/2302.12422v2", "titleZh": "MimicPlay：通过观察人类自由交互学习长时序模仿策略", "abstractZh": "从人类自由交互视频预测未来三维手部轨迹，学习目标条件的高层潜在计划；低层控制器用少量遥操作数据把计划映射成机器人动作。\n六种桌面环境、14项长时序任务；机器人示教为每任务20或40条，比较额外10分钟人类自由交互与基线额外机器人示教。", "summary": "从人类自由交互视频预测未来三维手部轨迹，学习目标条件的高层潜在计划；低层控制器用少量遥操作数据把计划映射成机器人动作。", "experimentType": "real", "robots": [ "Franka Emika机械臂（原文未细分型号）" ], "tags": [ "模仿学习", "操作与抓取" ], "limitations": "高层计划来自特定场景，主要是桌面任务；跨人机形态表示与移动操作仍待扩展。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的人类自由交互到机器人长时序学习研究。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2302.12422v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2302.12422v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/j96w/MimicPlay", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/j96w/MimicPlay", "category": "模仿学习 / 操作与抓取", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "训练、基线和评估脚本已发布。 作者提供 play 数据和视频提示下载链接。 作者提供仿真训练检查点链接。 公开仿真使用 robot play，同一批机器人数据训练两层，即 MimicPlay(0-human)；不能当作论文实机人类视频方案的完整复现。 本次未运行训练、复现实验或实机控制。", "directions": [ "模仿学习", "操作与抓取" ], "robotFilters": [ "Franka Emika机械臂（原文未细分型号）" ], "original": { "id": "arxiv-2302.12422", "originalSourceUrl": "https://arxiv.org/pdf/2302.12422v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2302.12422v2", "finalDownloadUrl": "https://arxiv.org/pdf/2302.12422v2", "sha256": "ba6056f66abae90f12c95cc44f9b6714cf6766a6ed7eec6f19766aa0f2685ee5", "bytes": 11211178, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:01:25.265828+00:00", "publicationDate": "2023/02/24", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2302.12422v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2302.12422v2", "pages": 21, "sourceIdentityStatus": "title_match", "sourceTitle": "MimicPlay: Long-Horizon Imitation Learning by Watching Human Play", "sourceVersion": "2302.12422v2", "sourceVersionDate": "2023/10/13", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2302.12422v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2302.12422" }, "originalAnalysis": { "id": "arxiv-2302.12422", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2302.12422v2", "sourceVersion": "2302.12422v2", "sectionsRead": [ "PDF第5页选段（提取文本L301–318）", "PDF第7页选段（提取文本L684–701）", "PDF第8页选段（提取文本L798–815）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "从人类自由交互视频预测未来三维手部轨迹，学习目标条件的高层潜在计划；低层控制器用少量遥操作数据把计划映射成机器人动作。", "experimentsZh": "六种桌面环境、14项长时序任务；机器人示教为每任务20或40条，比较额外10分钟人类自由交互与基线额外机器人示教。", "resultsZh": "加入人类自由交互数据显著改善长时序完成率；原文对无人体数据变体的提升超过23%，但不能据此推断任意视频都可迁移。", "limitationsZh": "高层计划来自特定场景，主要是桌面任务；跨人机形态表示与移动操作仍待扩展。", "reproductionZh": "匹配总采集时间、目标图像、手部轨迹提取和低层示教量，保留场景内与未见任务的区别。", "experimentType": "real", "robots": [ "Franka Emika机械臂（原文未细分型号）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "从人类自由交互视频预测未来三维手部轨迹，学习目标条件的高层潜在计划；低层控制器用少量遥操作数据把计划映射成机器人动作。" }, { "section": "原文实验与结果相关选段", "note": "六种桌面环境、14项长时序任务；机器人示教为每任务20或40条，比较额外10分钟人类自由交互与基线额外机器人示教。 加入人类自由交互数据显著改善长时序完成率；原文对无人体数据变体的提升超过23%，但不能据此推断任意视频都可迁移。" }, { "section": "局限与复现条件", "note": "高层计划来自特定场景，主要是桌面任务；跨人机形态表示与移动操作仍待扩展。" } ], "originalSha256": "ba6056f66abae90f12c95cc44f9b6714cf6766a6ed7eec6f19766aa0f2685ee5", "readingScopeSha256": "a09d0b59c6271a6f67446ec74b945deb451853c634aa138d0c57e5845c32fdda", "textExtractionSha256": "b6484c2ebc8cdd4d6eba2d7757f20d2a317f1b7843e5c6c56b1d02e6ad6e0f14" }, "experimentNote": "六种桌面环境、14项长时序任务；机器人示教为每任务20或40条，比较额外10分钟人类自由交互与基线额外机器人示教。", "contribution": "从人类自由交互视频预测未来三维手部轨迹，学习目标条件的高层潜在计划；低层控制器用少量遥操作数据把计划映射成机器人动作。", "whyUseful": "匹配总采集时间、目标图像、手部轨迹提取和低层示教量，保留场景内与未见任务的区别。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "仓库 MIT；外部手部检测器、数据与下载权重应分别核对发布条款。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2302.04659", "title": "ManiSkill2: A Unified Benchmark for Generalizable Manipulation Skills", "date": "2023-02-09", "url": "https://arxiv.org/abs/2302.04659v1", "titleZh": "ManiSkill2：通用且可泛化操作技能的统一基准", "abstractZh": "在SAPIEN上统一刚体与软体操作、视觉观测及控制器接口；用刚体和MPM双向耦合、异步渲染与共享渲染服务器提高采样效率。\n20类技能、逾2000物体与400万示教帧，比较规划、模仿与强化学习，并展示PickCube策略迁移和塑形动作的真实对照。", "summary": "在SAPIEN上统一刚体与软体操作、视觉观测及控制器接口；用刚体和MPM双向耦合、异步渲染与共享渲染服务器提高采样效率。", "experimentType": "both", "robots": [ "Franka Emika Panda（仿真）", "实机型号未核实" ], "tags": [ "数据集与基准", "操作与抓取", "强化学习" ], "limitations": "场景多样性、真实感及域随机化仍有限；本工作更侧重低层短时序技能，照片级渲染会降低速度。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "foundation", "tierNote": "编辑分类：通用操作仿真、控制器和示教的基础设施，适合用作统一训练和评估入口。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2302.04659v1", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2302.04659v1", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/mani-skill/ManiSkill/tree/v0.5.3", "note": "已核验官方代码与文档；未运行复现；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/mani-skill/ManiSkill/tree/v0.5.3", "category": "数据集与基准 / 操作与抓取", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "旧ManiSkill2仓库明确迁移，并指向v0.5.3原始代码快照；该快照许可已核实，勿用当前ManiSkill主干替代原基准。 本次未运行训练、复现实验或实机控制。", "directions": [ "数据集与基准", "操作与抓取", "强化学习" ], "robotFilters": [ "Franka Emika Panda（仿真）", "实机型号未核实" ], "original": { "id": "arxiv-2302.04659", "originalSourceUrl": "https://arxiv.org/pdf/2302.04659v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2302.04659v1", "finalDownloadUrl": "https://arxiv.org/pdf/2302.04659v1", "sha256": "89fb794345d860ee696ca2f4d7a75059b15352b615794eeadb0e6075426136f0", "bytes": 7635048, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:02.866027+00:00", "publicationDate": "2023/02/09", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2302.04659v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2302.04659v1", "pages": 30, "sourceIdentityStatus": "title_match", "sourceTitle": "ManiSkill2: A Unified Benchmark for Generalizable Manipulation Skills", "sourceVersion": "2302.04659v1", "sourceVersionDate": "2023/02/09", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2302.04659v1", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2302.04659" }, "originalAnalysis": { "id": "arxiv-2302.04659", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2302.04659v1", "sourceVersion": "2302.04659v1", "sectionsRead": [ "PDF第2页选段（提取文本L72–89）", "PDF第5页选段（提取文本L284–301）", "PDF第5–6页选段（提取文本L310–327）", "PDF第8页选段（提取文本L633–646）", "PDF第6页选段（提取文本L450–490）", "PDF第5页选段（提取文本L260–299）", "PDF第10–11页选段（提取文本L785–825）", "PDF第16页选段（提取文本L1090–1120）", "PDF第30页选段（提取文本L2100–2110）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "在SAPIEN上统一刚体与软体操作、视觉观测及控制器接口；用刚体和MPM双向耦合、异步渲染与共享渲染服务器提高采样效率。", "experimentsZh": "20类技能、逾2000物体与400万示教帧，比较规划、模仿与强化学习，并展示PickCube策略迁移和塑形动作的真实对照。", "resultsZh": "PickCube实机试验报告60%成功，深度域差距造成下降；软体塑形对照展示仿真潜力，不能推出所有任务均可直接迁移。", "limitationsZh": "场景多样性、真实感及域随机化仍有限；本工作更侧重低层短时序技能，照片级渲染会降低速度。", "reproductionZh": "须同时固定SAPIEN/ManiSkill2版本、控制模式、对象划分和渲染设置；后续ManiSkill版本不可直接代替原论文评估。", "experimentType": "both", "robots": [ "Franka Emika Panda（仿真）", "实机型号未核实" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "在SAPIEN上统一刚体与软体操作、视觉观测及控制器接口；用刚体和MPM双向耦合、异步渲染与共享渲染服务器提高采样效率。" }, { "section": "原文实验与结果相关选段", "note": "20类技能、逾2000物体与400万示教帧，比较规划、模仿与强化学习，并展示PickCube策略迁移和塑形动作的真实对照。 PickCube实机试验报告60%成功，深度域差距造成下降；软体塑形对照展示仿真潜力，不能推出所有任务均可直接迁移。" }, { "section": "局限与复现条件", "note": "场景多样性、真实感及域随机化仍有限；本工作更侧重低层短时序技能，照片级渲染会降低速度。" } ], "originalSha256": "89fb794345d860ee696ca2f4d7a75059b15352b615794eeadb0e6075426136f0", "readingScopeSha256": "a44560be12aa5525fa63436dcc84426c90ef3e1de44222162a1d7dc5fc079ca6", "textExtractionSha256": "ad86a9590b3558b1d26fdd4539f60018656fedafb1ee39b4f7d073fe1b225d61" }, "experimentNote": "20类技能、逾2000物体与400万示教帧，比较规划、模仿与强化学习，并展示PickCube策略迁移和塑形动作的真实对照。", "contribution": "在SAPIEN上统一刚体与软体操作、视觉观测及控制器接口；用刚体和MPM双向耦合、异步渲染与共享渲染服务器提高采样效率。", "whyUseful": "须同时固定SAPIEN/ManiSkill2版本、控制模式、对象划分和渲染设置；后续ManiSkill版本不可直接代替原论文评估。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "v0.5.3核心Apache-2.0；SAPIEN及物体数据独立许可。", "codeReleaseScope": "已核验官方代码与文档；未运行复现" }, { "id": "arxiv-2302.00111", "title": "Learning Universal Policies via Text-Guided Video Generation", "date": "2023-01-31", "url": "https://arxiv.org/abs/2302.00111v3", "titleZh": "UniPi：通过文本引导视频生成学习通用策略", "abstractZh": "把语言目标条件化的未来视频当作通用计划，再由任务相关逆动力学模型映射为动作；用分层时间超分辨生成更细致的轨迹。\n在模拟组合抓放及多任务环境执行策略；真实Bridge部分主要评估视频生成，并用末帧成功分类器作替代指标，没有据此证明实机闭环成功。", "summary": "把语言目标条件化的未来视频当作通用计划，再由任务相关逆动力学模型映射为动作；用分层时间超分辨生成更细致的轨迹。", "experimentType": "sim", "robots": [], "tags": [ "世界模型", "操作与抓取", "视觉语言动作" ], "limitations": "高保真视频生成可能需约一分钟；主要假设充分可观测，部分可观测时可能生成不存在的物体或运动。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的文本到视频规划研究，适合辨析视觉想象指标与真实执行能力。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2302.00111v3", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2302.00111v3", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "世界模型 / 操作与抓取", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "已核验原文；本轮未独立核实官方实现、训练入口及发布范围，不能据论文声明认定代码已开放。 本次未运行训练、复现实验或实机控制。", "directions": [ "世界模型", "操作与抓取", "视觉语言动作" ], "robotFilters": [], "original": { "id": "arxiv-2302.00111", "originalSourceUrl": "https://arxiv.org/pdf/2302.00111v3", "versionedOriginalUrl": "https://arxiv.org/pdf/2302.00111v3", "finalDownloadUrl": "https://arxiv.org/pdf/2302.00111v3", "sha256": "5ded462cfcad30a6317610a4f991d3c1eca7826cbe208dbd9929c3837529708e", "bytes": 6171147, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:58.877825+00:00", "publicationDate": "2023/01/31", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2302.00111v3", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2302.00111v3", "pages": 17, "sourceIdentityStatus": "title_match", "sourceTitle": "Learning Universal Policies via Text-Guided Video Generation", "sourceVersion": "2302.00111v3", "sourceVersionDate": "2023/11/20", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2302.00111v3", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2302.00111" }, "originalAnalysis": { "id": "arxiv-2302.00111", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2302.00111v3", "sourceVersion": "2302.00111v3", "sectionsRead": [ "PDF第5页选段（提取文本L332–349）", "PDF第9页选段（提取文本L682–699）", "PDF第10页选段（提取文本L747–764）", "PDF第3页选段（提取文本L144–178）", "PDF第5页选段（提取文本L292–326）", "PDF第8–9页选段（提取文本L631–685）", "PDF第8页选段（提取文本L608–630）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "把语言目标条件化的未来视频当作通用计划，再由任务相关逆动力学模型映射为动作；用分层时间超分辨生成更细致的轨迹。", "experimentsZh": "在模拟组合抓放及多任务环境执行策略；真实Bridge部分主要评估视频生成，并用末帧成功分类器作替代指标，没有据此证明实机闭环成功。", "resultsZh": "预训练改善视频计划与指令组合泛化；Bridge表中的77.1%是生成视频分类指标，不能标为机器人操作成功率。", "limitationsZh": "高保真视频生成可能需约一分钟；主要假设充分可观测，部分可观测时可能生成不存在的物体或运动。", "reproductionZh": "分别核对视频质量、分类器替代分数和实际动作执行，记录逆动力学训练数据与扩散采样开销。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "把语言目标条件化的未来视频当作通用计划，再由任务相关逆动力学模型映射为动作；用分层时间超分辨生成更细致的轨迹。" }, { "section": "原文实验与结果相关选段", "note": "在模拟组合抓放及多任务环境执行策略；真实Bridge部分主要评估视频生成，并用末帧成功分类器作替代指标，没有据此证明实机闭环成功。 预训练改善视频计划与指令组合泛化；Bridge表中的77.1%是生成视频分类指标，不能标为机器人操作成功率。" }, { "section": "局限与复现条件", "note": "高保真视频生成可能需约一分钟；主要假设充分可观测，部分可观测时可能生成不存在的物体或运动。" } ], "originalSha256": "5ded462cfcad30a6317610a4f991d3c1eca7826cbe208dbd9929c3837529708e", "readingScopeSha256": "b1ae79e3f6796745c75b6ba1fedb48f65e1bf17ae6568ce9b31b286fefaf083f", "textExtractionSha256": "95cccd55bf6e0a580b7924b2fd283bb4db44900dba1f433441aa0ad57dc3046f" }, "experimentNote": "在模拟组合抓放及多任务环境执行策略；真实Bridge部分主要评估视频生成，并用末帧成功分类器作替代指标，没有据此证明实机闭环成功。", "contribution": "把语言目标条件化的未来视频当作通用计划，再由任务相关逆动力学模型映射为动作；用分层时间超分辨生成更细致的轨迹。", "whyUseful": "分别核对视频质量、分类器替代分数和实际动作执行，记录逆动力学训练数据与扩散采样开销。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": null, "codeReleaseScope": "未独立核实" }, { "id": "arxiv-2301.10602", "title": "DreamWaQ: Learning Robust Quadrupedal Locomotion With Implicit Terrain Imagination via Deep Reinforcement Learning", "date": "2023-01-25", "url": "https://arxiv.org/abs/2301.10602v2", "titleZh": "DreamWaQ：通过隐式地形想象学习鲁棒四足运动", "abstractZh": "上下文辅助估计网络联合学习机体速度与潜在环境表示，用本体感觉历史支撑RL控制；奖励兼顾运动表现及电机功率分布。\n仿真以十个随机种子比较估计器与策略；实机A1上估计与策略50Hz、关节PD200Hz，包含滑移、绊脚及430米和465米户外路线。", "summary": "上下文辅助估计网络联合学习机体速度与潜在环境表示，用本体感觉历史支撑RL控制；奖励兼顾运动表现及电机功率分布。", "experimentType": "both", "robots": [ "Unitree A1" ], "tags": [ "运动控制", "强化学习" ], "limitations": "评测范围限制：仅本体感觉难以直接提前观测远处障碍；实机控制增益、载荷和接触条件仍影响迁移，未证明任意障碍安全。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的本体感觉运动适应研究，关注状态估计与潜在上下文。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2301.10602v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2301.10602v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "运动控制 / 强化学习", "year": 2023, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "已核验原文；本轮未独立核实官方实现、训练入口及发布范围，不能据论文声明认定代码已开放。 本次未运行训练、复现实验或实机控制。", "directions": [ "运动控制", "强化学习" ], "robotFilters": [ "Unitree A1" ], "original": { "id": "arxiv-2301.10602", "originalSourceUrl": "https://arxiv.org/pdf/2301.10602v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2301.10602v2", "finalDownloadUrl": "https://arxiv.org/pdf/2301.10602v2", "sha256": "fb02d20879c38b2ca7d0b52ff58174f2119451dea23e946c3eecfccdfcce1cfd", "bytes": 2788181, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:31.157800+00:00", "publicationDate": "2023/01/25", "license": "http://creativecommons.org/licenses/by-nc-nd/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2301.10602v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2301.10602v2", "pages": 7, "sourceIdentityStatus": "title_match", "sourceTitle": "DreamWaQ: Learning Robust Quadrupedal Locomotion With Implicit Terrain Imagination via Deep Reinforcement Learning", "sourceVersion": "2301.10602v2", "sourceVersionDate": "2023/03/03", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2301.10602v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2301.10602" }, "originalAnalysis": { "id": "arxiv-2301.10602", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2301.10602v2", "sourceVersion": "2301.10602v2", "sectionsRead": [ "PDF第3页选段（提取文本L318–335）", "PDF第4页选段（提取文本L500–513）", "PDF第4页选段（提取文本L517–534）", "PDF第3页选段（提取文本L334–374）", "PDF第5–6页选段（提取文本L634–665）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "上下文辅助估计网络联合学习机体速度与潜在环境表示，用本体感觉历史支撑RL控制；奖励兼顾运动表现及电机功率分布。", "experimentsZh": "仿真以十个随机种子比较估计器与策略；实机A1上估计与策略50Hz、关节PD200Hz，包含滑移、绊脚及430米和465米户外路线。", "resultsZh": "学习的状态与上下文估计支持不依赖地形视觉的行走适应；长距离测试是具体路线实证，不能泛化为所有地形可靠性。", "limitationsZh": "评测范围限制：仅本体感觉难以直接提前观测远处障碍；实机控制增益、载荷和接触条件仍影响迁移，未证明任意障碍安全。", "reproductionZh": "保留CENet历史窗口、速度估计训练、地形课程、域随机化和PD增益；不要把训练GPU耗时等同实机学习时间。", "experimentType": "both", "robots": [ "Unitree A1" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "上下文辅助估计网络联合学习机体速度与潜在环境表示，用本体感觉历史支撑RL控制；奖励兼顾运动表现及电机功率分布。" }, { "section": "原文实验与结果相关选段", "note": "仿真以十个随机种子比较估计器与策略；实机A1上估计与策略50Hz、关节PD200Hz，包含滑移、绊脚及430米和465米户外路线。 学习的状态与上下文估计支持不依赖地形视觉的行走适应；长距离测试是具体路线实证，不能泛化为所有地形可靠性。" }, { "section": "局限与复现条件", "note": "评测范围限制：仅本体感觉难以直接提前观测远处障碍；实机控制增益、载荷和接触条件仍影响迁移，未证明任意障碍安全。" } ], "originalSha256": "fb02d20879c38b2ca7d0b52ff58174f2119451dea23e946c3eecfccdfcce1cfd", "readingScopeSha256": "17fbbda2d09ee8617bb08ec252fa4afd4332b90236bc9d5931dbad67a07a6058", "textExtractionSha256": "7d4265bb71dbf21c33e4a5c7a64d333959986ba5b4dfbb023e607a48d4c827c7" }, "experimentNote": "仿真以十个随机种子比较估计器与策略；实机A1上估计与策略50Hz、关节PD200Hz，包含滑移、绊脚及430米和465米户外路线。", "contribution": "上下文辅助估计网络联合学习机体速度与潜在环境表示，用本体感觉历史支撑RL控制；奖励兼顾运动表现及电机功率分布。", "whyUseful": "保留CENet历史窗口、速度估计训练、地形课程、域随机化和PD增益；不要把训练GPU耗时等同实机学习时间。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": null, "codeReleaseScope": "未独立核实" }, { "id": "arxiv-2301.04104", "title": "Mastering Diverse Domains through World Models", "titleZh": "通过世界模型掌握多样化任务", "shortTitle": "DreamerV3", "date": "2023-01-10", "year": 2023, "url": "https://arxiv.org/abs/2301.04104", "codeUrl": "https://github.com/danijar/dreamerv3", "category": "世界模型与强化学习", "tags": [ "世界模型", "潜空间想象", "强化学习", "仿真" ], "tier": "foundation", "summary": "DreamerV3以离散随机潜变量和循环状态学习可重构观测、预测奖励及终止的世界模型，再在想象轨迹中训练演员和评论家。symlog/双热分类处理跨量级信号，KL平衡、自由比特与均匀概率混合稳定训练；执行时直接采…", "abstractZh": "DreamerV3以离散随机潜变量和循环状态学习可重构观测、预测奖励及终止的世界模型，再在想象轨迹中训练演员和评论家。symlog/双热分类处理跨量级信号，KL平衡、自由比特与均匀概率混合稳定训练；执行时直接采样演员，不做在线树搜索。\n原文版本在八类、150余任务评测固定超参数，覆盖Atari、ProcGen、DMLab、连续控制、BSuite及Minecraft；不同任务分别训练智能体。另在14任务消融学习信号，并比较12M至400M参数及经验回放比例，每智能体使用单A100。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "sim", "robots": [ "DeepMind Control Suite 虚拟控制体", "Minecraft、Atari、DMLab 等游戏智能体（非实机机器人）" ], "codeStatus": "open", "status": "作者维护 MIT 实现；当前仓库标注为重实现。", "trainingNote": "在线强化学习需要环境交互与奖励；多个领域通常分别训练，不是把同一权重零样本部署到所有任务。", "whyUseful": "应核对所用论文/代码版本、环境动作重复、终止规则、奖励变换与回放比，并区分帧数和环境步。先复现小型控制任务和损失稳定性，再扩大规模，本次未运行训练。", "contribution": "DreamerV3以离散随机潜变量和循环状态学习可重构观测、预测奖励及终止的世界模型，再在想象轨迹中训练演员和评论家。symlog/双热分类处理跨量级信号，KL平衡、自由比特与均匀概率混合稳定训练；执行时直接采样演员，不做在线树搜索。", "limitations": "共享的是算法和超参数，并非一个检查点掌握全部领域；全部是游戏或模拟控制，无实机证据。学习仍依赖大量交互及奖励，跨基准对照预算、环境设定不同；规模化节省交互不等于减少总算力。", "caveats": "2023 为预印本首发；2025 Nature 发表题名为 Mastering diverse control tasks through world models。", "evidence": [ { "url": "https://arxiv.org/abs/2301.04104", "note": "2023-01-10 首发与研究摘要。" }, { "url": "https://arxiv.org/pdf/2301.04104", "note": "实验基准为控制仿真与游戏；无原文实机评测。" }, { "url": "https://github.com/danijar/dreamerv3", "note": "作者 MIT 重实现及 2025 正式发表引用。" }, { "url": "https://arxiv.org/pdf/2301.04104", "note": "pp3–7：RSSM、想象演员评论家与稳定损失。" }, { "url": "https://arxiv.org/pdf/2301.04104", "note": "pp8–10; Figure 6：八域150余任务、各任务独立训练、Minecraft一亿步及规模消融。" } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "世界模型", "强化学习" ], "robotFilters": [], "verificationNote": "已核预印本、更新稿实验与官方代码；不把仿真控制归为实机训练。", "original": { "id": "arxiv-2301.04104", "originalSourceUrl": "https://arxiv.org/pdf/2301.04104v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2301.04104v2", "finalDownloadUrl": "https://arxiv.org/pdf/2301.04104", "sha256": "d0385798e8bada8e81b915c1743be81d9ce8776f12ed937e09351995e099ce37", "bytes": 2874717, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:40.867947+00:00", "publicationDate": "2023/01/10", "publicationDateSourceUrl": "https://arxiv.org/abs/2301.04104", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2301.04104", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2301.04104", "pages": 40, "sourceIdentityStatus": "title_match", "sourceTitle": "Mastering Diverse Domains through World Models", "sourceVersion": "2301.04104v2", "sourceVersionDate": "2024/04/17", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.169273+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2301.04104" }, "originalAnalysis": { "id": "arxiv-2301.04104", "sourceUrl": "https://arxiv.org/pdf/2301.04104", "sectionsRead": [ "Learning algorithm", "World model/Critic learning", "Robust predictions", "Results pp8–10", "Figure 6" ], "methodsZh": "DreamerV3以离散随机潜变量和循环状态学习可重构观测、预测奖励及终止的世界模型，再在想象轨迹中训练演员和评论家。symlog/双热分类处理跨量级信号，KL平衡、自由比特与均匀概率混合稳定训练；执行时直接采样演员，不做在线树搜索。", "experimentsZh": "原文版本在八类、150余任务评测固定超参数，覆盖Atari、ProcGen、DMLab、连续控制、BSuite及Minecraft；不同任务分别训练智能体。另在14任务消融学习信号，并比较12M至400M参数及经验回放比例，每智能体使用单A100。", "resultsZh": "作者报告在多类基准优于固定超参数PPO，并在Minecraft各受测种子中于一亿环境步内找到钻石。重构学习信号对成绩贡献明显，模型更大和更多更新常降低交互需求；此处“找到钻石”不是零训练或单回合必成。", "limitationsZh": "共享的是算法和超参数，并非一个检查点掌握全部领域；全部是游戏或模拟控制，无实机证据。学习仍依赖大量交互及奖励，跨基准对照预算、环境设定不同；规模化节省交互不等于减少总算力。", "reproductionZh": "应核对所用论文/代码版本、环境动作重复、终止规则、奖励变换与回放比，并区分帧数和环境步。先复现小型控制任务和损失稳定性，再扩大规模，本次未运行训练。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "pp3–7", "note": "RSSM、想象演员评论家与稳定损失。" }, { "section/page": "pp8–10; Figure 6", "note": "八域150余任务、各任务独立训练、Minecraft一亿步及规模消融。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "原文版本在八类、150余任务评测固定超参数，覆盖Atari、ProcGen、DMLab、连续控制、BSuite及Minecraft；不同任务分别训练智能体。另在14任务消融学习信号，并比较12M至400M参数及经验回放比例，每智能体使用单A100。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2212.06817", "title": "RT-1: Robotics Transformer for Real-World Control at Scale", "titleZh": "RT-1：面向大规模真实世界控制的机器人 Transformer", "shortTitle": "RT-1", "date": "2022-12-13", "year": 2022, "url": "https://arxiv.org/abs/2212.06817", "codeUrl": "https://github.com/google-research/robotics_transformer", "category": "机器人基础模型", "tags": [ "Transformer 架构", "模仿学习", "多任务", "实机" ], "tier": "foundation", "summary": "RT-1用语言嵌入FiLM调制EfficientNet图像特征，TokenLearner把每帧压成8词元，六帧历史送入Transformer。机械臂、底盘与终止模式离散成动作词元，以监督交叉熵学习；压缩及非逐t…", "abstractZh": "RT-1用语言嵌入FiLM调制EfficientNet图像特征，TokenLearner把每帧压成8词元，六帧历史送入Transformer。机械臂、底盘与终止模式离散成动作词元，以监督交叉熵学习；压缩及非逐token自回归等优化满足约3Hz真机控制。\n13台Everyday Robots移动机械臂、17个月约13万遥操作轨迹，覆盖744条任务指令，约3000次真实评估；分已见、未见组合、干扰物及新背景。再混入仿真和Kuka数据检验数据兼容，结合SayCan做长任务。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "real", "robots": [ "Everyday Robots移动机械臂" ], "codeStatus": "open", "status": "官方实现、检查点和数据已公开；代码仓库已归档。", "trainingNote": "Apache-2.0；13 台实机收集约 13 万轨迹。含 KUKA 与仿真数据混训实验，主评测仍在 Everyday Robots 实机。", "whyUseful": "先固定六帧历史、每维256bins、相机预处理和实际延迟预算；按任务/对象/场景而非逐帧划分数据，重现数据量与多样性消融。应分别报告RT-1单技能和接SayCan后的长程成功，不归为同一端到端模型能力。", "contribution": "RT-1用语言嵌入FiLM调制EfficientNet图像特征，TokenLearner把每帧压成8词元，六帧历史送入Transformer。机械臂、底盘与终止模式离散成动作词元，以监督交叉熵学习；压缩及非逐token自回归等优化满足约3Hz真机控制。", "limitations": "所谓744技能很多是对象与已有动作的语言组合，未验证任意新运动。行为克隆不能自然超过示范者，任务整体精巧程度有限；大规模采集成本很高，背景变化下仍显著下降。", "caveats": "KUKA 是迁移训练数据来源，不能据此认定 RT-1 在 KUKA 上完成实机评测。", "evidence": [ { "url": "https://arxiv.org/abs/2212.06817", "note": "首发日期与摘要。" }, { "url": "https://arxiv.org/html/2212.06817v1", "note": "第 6 节及附录明确 Everyday Robots 实机评测与跨机器人数据实验。" }, { "url": "https://github.com/google-research/robotics_transformer", "note": "Apache-2.0、已发布检查点及归档状态。" }, { "url": "https://arxiv.org/pdf/2212.06817", "note": "5.1–5.2; Table 1：词元化、3Hz设计及744指令数据集。" }, { "url": "https://arxiv.org/pdf/2212.06817", "note": "6 Tables 2–7; 7：分项成绩、异构数据与组合泛化边界。" }, { "url": "https://arxiv.org/pdf/2212.06817v2", "note": "5.1–5.2; Table 1：词元化、3Hz设计及744指令数据集。" }, { "url": "https://arxiv.org/pdf/2212.06817v2", "note": "6 Tables 2–7; 7：分项成绩、异构数据与组合泛化边界。" } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "视觉语言动作", "模仿学习" ], "robotFilters": [], "verificationNote": "已核论文、官方项目与代码；实验类别按评测而非训练数据来源标注。", "original": { "id": "arxiv-2212.06817", "originalSourceUrl": "https://arxiv.org/pdf/2212.06817v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2212.06817v2", "finalDownloadUrl": "https://arxiv.org/pdf/2212.06817", "sha256": "81efe669c8fa50ffc7097886dc78a442fa7a8f7f2451684b93069cb9fbeda767", "bytes": 14672588, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:49.019582+00:00", "publicationDate": "2022/12/13", "publicationDateSourceUrl": "https://arxiv.org/abs/2212.06817", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2212.06817", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2212.06817", "pages": 31, "sourceIdentityStatus": "title_match", "sourceTitle": "RT-1: Robotics Transformer for Real-World Control at Scale", "sourceVersion": "2212.06817v2", "sourceVersionDate": "2023/08/11", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.180896+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2212.06817" }, "originalAnalysis": { "id": "arxiv-2212.06817", "sourceUrl": "https://arxiv.org/pdf/2212.06817v2", "sectionsRead": [ "5.1–5.2", "6.1–6.5", "7", "Tables 1–7" ], "methodsZh": "RT-1用语言嵌入FiLM调制EfficientNet图像特征，TokenLearner把每帧压成8词元，六帧历史送入Transformer。机械臂、底盘与终止模式离散成动作词元，以监督交叉熵学习；压缩及非逐token自回归等优化满足约3Hz真机控制。", "experimentsZh": "13台Everyday Robots移动机械臂、17个月约13万遥操作轨迹，覆盖744条任务指令，约3000次真实评估；分已见、未见组合、干扰物及新背景。再混入仿真和Kuka数据检验数据兼容，结合SayCan做长任务。", "resultsZh": "主表已见97%、未见组合76%、干扰物83%、背景59%；同是语言条件的BC-Z/Gato表现更低。机器人异构数据能补充技能；数据消融表明保留任务多样性比只保留更多同类示范更关键。长流程结果依赖SayCan的上层技能编排。", "limitationsZh": "所谓744技能很多是对象与已有动作的语言组合，未验证任意新运动。行为克隆不能自然超过示范者，任务整体精巧程度有限；大规模采集成本很高，背景变化下仍显著下降。", "reproductionZh": "先固定六帧历史、每维256bins、相机预处理和实际延迟预算；按任务/对象/场景而非逐帧划分数据，重现数据量与多样性消融。应分别报告RT-1单技能和接SayCan后的长程成功，不归为同一端到端模型能力。", "experimentType": "real", "robots": [ "Everyday Robots移动机械臂" ], "corrections": [ "主要控制实验证据为实机；加入仿真训练数据不等于主成功率来自仿真。" ], "evidenceNotes": [ { "note": "词元化、3Hz设计及744指令数据集。", "section": "5.1–5.2; Table 1" }, { "note": "分项成绩、异构数据与组合泛化边界。", "section": "6 Tables 2–7; 7" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2212.06817v2", "originalSha256": "81efe669c8fa50ffc7097886dc78a442fa7a8f7f2451684b93069cb9fbeda767" }, "experimentNote": "13台Everyday Robots移动机械臂、17个月约13万遥操作轨迹，覆盖744条任务指令，约3000次真实评估；分已见、未见组合、干扰物及新背景。再混入仿真和Kuka数据检验数据兼容，结合SayCan做长任务。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2211.09423", "title": "DexPoint: Generalizable Point Cloud Reinforcement Learning for Sim-to-Real Dexterous Manipulation", "titleZh": "DexPoint：面向仿真到现实灵巧操作的可泛化点云强化学习", "shortTitle": "DexPoint", "date": "2022-11-17", "datePrecision": "day", "year": 2022, "url": "https://arxiv.org/abs/2211.09423", "paperUrl": "https://arxiv.org/abs/2211.09423", "codeUrl": "https://github.com/yzqin/dexpoint-release", "summary": "DexPoint将深度相机点云与依据手部网格、关节编码器正运动学合成的“想象手点云”融合，补全遮挡手指；以PointNet类特征联合本体状态、目标位置驱动PPO。训练奖励利用仿真接触真值鼓励拇指与其他手指共同包…", "abstractZh": "DexPoint将深度相机点云与依据手部网格、关节编码器正运动学合成的“想象手点云”融合，补全遮挡手指；以PointNet类特征联合本体状态、目标位置驱动PPO。训练奖励利用仿真接触真值鼓励拇指与其他手指共同包围物体，但部署观测不需要接触真值。\n在抓取搬运和转动门把手开门两类任务比较单物体、多物体训练及移除接触奖励/想象点云的消融；仿真五随机种子并做100次评估。实机为xArm6搭配Allegro手，测试26个抓取物体、三扇门，每物体与策略组合进行10次独立试验。", "category": "灵巧手 / 三维视觉控制", "tags": [ "点云", "Sim2Real", "接触奖励", "类别泛化" ], "directions": [ "灵巧手", "操作与抓取", "强化学习" ], "tier": "classic", "experimentType": "both", "robots": [ "xArm6", "Allegro Hand" ], "robotNote": "", "codeStatus": "open", "status": "作者仿真环境公开，并链接共享的PPO训练代码。", "trainingNote": "在SAPIEN点云环境中训练PPO；官方仓库说明训练接口复用DexArt代码，并需下载对象资产。", "contribution": "DexPoint将深度相机点云与依据手部网格、关节编码器正运动学合成的“想象手点云”融合，补全遮挡手指；以PointNet类特征联合本体状态、目标位置驱动PPO。训练奖励利用仿真接触真值鼓励拇指与其他手指共同包围物体，但部署观测不需要接触真值。", "whyUseful": "复现须工作区裁剪、512点采样、距离相关深度噪声、手眼标定和22维手臂动作映射。保持仿真/实机预处理一致并报告接触奖励权重；零实机微调不等于无需标定和人工目标设置。", "limitations": "只覆盖两种任务和有限类别，未证明广泛手内操作；无时序策略和触觉，严重遮挡仍会破坏点云特征。真实抓取以平面目标误差5厘米且抬高15厘米为成功，不能直接与仿真三维5厘米阈值混算。", "evidence": [ { "url": "https://arxiv.org/abs/2211.09423", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/pdf/2211.09423", "note": "§4.4明确Allegro + XArm-6实机；§5列出仅两任务的限制。" }, { "url": "https://yzqin.github.io/dexpoint/", "note": "官方方法介绍及实机失败案例。" }, { "url": "https://github.com/yzqin/dexpoint-release", "note": "README明确环境、资产与复用DexArt的PPO训练入口。" }, { "url": "https://arxiv.org/pdf/2211.09423v2", "note": "3.1–3.3：Contact reward is privileged training signal; imagined hand cloud uses forward kinematics, available at deployment." }, { "url": "https://arxiv.org/pdf/2211.09423v2", "note": "4.1：512 points, hand-eye calibration; different real/simulation success criteria." }, { "url": "https://arxiv.org/pdf/2211.09423v2", "note": "4.4 表3/4：xArm-6+Allegro; 26 objects; 10 trials/object-policy; 0.87/0.83/0.73 grasp success." }, { "url": "https://arxiv.org/pdf/2211.09423v2", "note": "5：Authors limit experiments to two tasks and propose recurrent temporal information as future work." } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2022：几何观测与身体先验支持灵巧手策略直接跨越仿真和真实对象。", "projectUrl": null, "robotFilters": [ "xArm6", "Allegro Hand" ], "original": { "id": "arxiv-2211.09423", "originalSourceUrl": "https://arxiv.org/pdf/2211.09423v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2211.09423v2", "finalDownloadUrl": "https://arxiv.org/pdf/2211.09423", "sha256": "c77516db0f09ad6c3964ee279b56c4db445502827af860fa3c1f5c7c517ae93d", "bytes": 6899060, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:52.686379+00:00", "publicationDate": "2022/11/17", "publicationDateSourceUrl": "https://arxiv.org/abs/2211.09423", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2211.09423", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2211.09423", "pages": 13, "sourceIdentityStatus": "title_match", "sourceTitle": "DexPoint: Generalizable Point Cloud Reinforcement Learning for Sim-to-Real Dexterous Manipulation", "sourceVersion": "2211.09423v2", "sourceVersionDate": "2022/11/18", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.187062+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2211.09423" }, "originalAnalysis": { "id": "arxiv-2211.09423", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739391+00:00", "sourceUrl": "https://arxiv.org/pdf/2211.09423v2", "sectionsRead": [ "3 任务、观测与动作", "3.1–3.3 接触奖励、想象点云与PPO", "4.1–4.4 实验协议及表1–4", "5 局限" ], "methodsZh": "DexPoint将深度相机点云与依据手部网格、关节编码器正运动学合成的“想象手点云”融合，补全遮挡手指；以PointNet类特征联合本体状态、目标位置驱动PPO。训练奖励利用仿真接触真值鼓励拇指与其他手指共同包围物体，但部署观测不需要接触真值。", "experimentsZh": "在抓取搬运和转动门把手开门两类任务比较单物体、多物体训练及移除接触奖励/想象点云的消融；仿真五随机种子并做100次评估。实机为xArm6搭配Allegro手，测试26个抓取物体、三扇门，每物体与策略组合进行10次独立试验。", "resultsZh": "作者报告多物体训练实机瓶、罐、混合类成功率为87%、83%、73%，两扇新门为60%、67%。去掉接触奖励后仿真抓取接近失败，想象点云对瓶类协调和遮挡门把手尤有帮助；对照EigenGrasp还需物体模型或位姿。", "limitationsZh": "只覆盖两种任务和有限类别，未证明广泛手内操作；无时序策略和触觉，严重遮挡仍会破坏点云特征。真实抓取以平面目标误差5厘米且抬高15厘米为成功，不能直接与仿真三维5厘米阈值混算。", "reproductionZh": "复现须工作区裁剪、512点采样、距离相关深度噪声、手眼标定和22维手臂动作映射。保持仿真/实机预处理一致并报告接触奖励权重；零实机微调不等于无需标定和人工目标设置。", "evidenceNotes": [ { "section/page": "3.1–3.3", "note": "Contact reward is privileged training signal; imagined hand cloud uses forward kinematics, available at deployment." }, { "section/page": "4.1", "note": "512 points, hand-eye calibration; different real/simulation success criteria." }, { "section/page": "4.4 表3/4", "note": "xArm-6+Allegro; 26 objects; 10 trials/object-policy; 0.87/0.83/0.73 grasp success." }, { "section/page": "5", "note": "Authors limit experiments to two tasks and propose recurrent temporal information as future work." } ], "experimentType": "both", "robots": [ "xArm6", "Allegro Hand" ], "corrections": [ "区分训练用接触真值与部署输入；不能称部署依赖触觉传感器。" ], "sourceVersion": "2211.09423v2", "originalSha256": "c77516db0f09ad6c3964ee279b56c4db445502827af860fa3c1f5c7c517ae93d" }, "experimentNote": "在抓取搬运和转动门把手开门两类任务比较单物体、多物体训练及移除接触奖励/想象点云的消融；仿真五随机种子并做100次评估。实机为xArm6搭配Allegro手，测试26个抓取物体、三扇门，每物体与策略组合进行10次独立试验。", "analysisVerifiedAt": "2026-10-04T13:47:58.739391+00:00" }, { "id": "arxiv-2210.10044", "title": "Deep Whole-Body Control: Learning a Unified Policy for Manipulation and Locomotion", "titleZh": "深度全身控制：学习统一的操作与移动策略", "shortTitle": "Deep Whole-Body Control", "date": "2022-10-18", "datePrecision": "day", "year": 2022, "url": "https://arxiv.org/abs/2210.10044", "paperUrl": "https://arxiv.org/abs/2210.10044", "codeUrl": "https://github.com/MarkFzp/Deep-Whole-Body-Control", "summary": "用单一强化学习策略同时输出12个腿与6个臂关节目标，协调底盘速度和末端目标。优势混合课程先分配臂/腿任务信用，再逐步耦合；正则化在线适应模块从状态历史估计环境潜变量，减少仿真特权编码到部署估计的差异。", "abstractZh": "用单一强化学习策略同时输出12个腿与6个臂关节目标，协调底盘速度和末端目标。优势混合课程先分配臂/腿任务信用，再逐步耦合；正则化在线适应模块从状态历史估计环境潜变量，减少仿真特权编码到部署估计的差异。\n实体平台是Unitree Go1加Interbotix WidowX250s和夹爪；仿真以三网络种子、各1000回合测试工作空间、扰动、跟踪和能耗。实机比较Go1内置MPC加臂IK，遥操作和AprilTag视觉反馈指定末端目标。", "category": "运动控制 / 足式移动操作", "tags": [ "全身控制", "移动操作", "在线适应", "优势混合" ], "directions": [ "运动控制", "操作与抓取" ], "tier": "classic", "experimentType": "both", "robots": [ "Unitree Go1", "Interbotix WidowX 250s" ], "robotNote": "", "codeStatus": "open", "status": "作者PyTorch参考实现公开。", "trainingNote": "Isaac Gym训练协调策略与适应模块；作者代码含legged_gym、rsl_rl和widowGo1模块，硬件需单独配置。", "contribution": "用单一强化学习策略同时输出12个腿与6个臂关节目标，协调底盘速度和末端目标。优势混合课程先分配臂/腿任务信用，再逐步耦合；正则化在线适应模块从状态历史估计环境潜变量，减少仿真特权编码到部署估计的差异。", "whyUseful": "复现需组合机器人惯性、电机、控制频率、环境随机化和优势混合调度，核对附录的机载分工。零实机微调只指已训练策略转移；还需基准MPC/IK和标记视觉设置，未把RaspberryPi控制计算等同通用视觉推理。", "limitations": "这是统一运动控制而非端到端语义操作；视觉目标来自AprilTag位置反馈，夹爪开合由人或脚本另控。物体交互仅初步，遮挡、软物体和更复杂接触仍待研究，小样本抓取不构成普适80%保证。", "evidence": [ { "url": "https://arxiv.org/abs/2210.10044", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/pdf/2210.10044", "note": "§3.1及附录F明确Go1、WidowX 250s、AprilTag和夹爪不在策略内。" }, { "url": "https://github.com/MarkFzp/Deep-Whole-Body-Control", "note": "作者仓库明确对应CoRL 2022论文的PyTorch实现。" }, { "url": "https://arxiv.org/pdf/2210.10044", "note": "§3.1：Go1+WidowX250s，机载供电计算。" }, { "url": "https://arxiv.org/pdf/2210.10044", "note": "Table 7：实机每设置10次、成功0.8，AprilTag引导。" }, { "url": "https://arxiv.org/pdf/2210.10044", "note": "Supplementary：夹爪开合不属于学得策略，遥控或脚本处理。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2022：将腿臂分离控制推进为可实机部署的统一足式移动操作策略。", "projectUrl": null, "robotFilters": [ "Unitree Go1", "Interbotix WidowX 250s" ], "original": { "id": "arxiv-2210.10044", "originalSourceUrl": "https://arxiv.org/pdf/2210.10044v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2210.10044v1", "finalDownloadUrl": "https://arxiv.org/pdf/2210.10044", "sha256": "836a835e9d8939c6261b3ca597a19bbafa3540dbde722c651c2dbe035f1f1541", "bytes": 7922418, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:38:53.010434+00:00", "publicationDate": "2022/10/18", "publicationDateSourceUrl": "https://arxiv.org/abs/2210.10044", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2210.10044", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2210.10044", "pages": 16, "sourceIdentityStatus": "title_match", "sourceTitle": "Deep Whole-Body Control: Learning a Unified Policy for Manipulation and Locomotion", "sourceVersion": "2210.10044v1", "sourceVersionDate": "2022/10/18", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.193366+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2210.10044" }, "originalAnalysis": { "id": "arxiv-2210.10044", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090141+00:00", "sourceUrl": "https://arxiv.org/pdf/2210.10044", "sectionsRead": [ "§2", "§3.1–3.3", "Tables 5–7", "§5", "Supplementary real-world details" ], "methodsZh": "用单一强化学习策略同时输出12个腿与6个臂关节目标，协调底盘速度和末端目标。优势混合课程先分配臂/腿任务信用，再逐步耦合；正则化在线适应模块从状态历史估计环境潜变量，减少仿真特权编码到部署估计的差异。", "experimentsZh": "实体平台是Unitree Go1加Interbotix WidowX250s和夹爪；仿真以三网络种子、各1000回合测试工作空间、扰动、跟踪和能耗。实机比较Go1内置MPC加臂IK，遥操作和AprilTag视觉反馈指定末端目标。", "resultsZh": "作者报告仿真臂工作空间0.82立方米，对分离策略0.58；扰动生存0.87对0.64。真实简单/困难抓取均80%，MPC+IK为30%/10%，完成时间约5/5.6秒；表格基于每设置10次真实试验。", "limitationsZh": "这是统一运动控制而非端到端语义操作；视觉目标来自AprilTag位置反馈，夹爪开合由人或脚本另控。物体交互仅初步，遮挡、软物体和更复杂接触仍待研究，小样本抓取不构成普适80%保证。", "reproductionZh": "复现需组合机器人惯性、电机、控制频率、环境随机化和优势混合调度，核对附录的机载分工。零实机微调只指已训练策略转移；还需基准MPC/IK和标记视觉设置，未把RaspberryPi控制计算等同通用视觉推理。", "experimentType": "both", "robots": [ "Unitree Go1", "Interbotix WidowX 250s" ], "evidenceNotes": [ { "section/page": "§3.1", "note": "Go1+WidowX250s，机载供电计算。" }, { "section/page": "Table 7", "note": "实机每设置10次、成功0.8，AprilTag引导。" }, { "section/page": "Supplementary", "note": "夹爪开合不属于学得策略，遥控或脚本处理。" } ] }, "experimentNote": "实体平台是Unitree Go1加Interbotix WidowX250s和夹爪；仿真以三网络种子、各1000回合测试工作空间、扰动、跟踪和能耗。实机比较Go1内置MPC加臂IK，遥操作和AprilTag视觉反馈指定末端目标。", "analysisVerifiedAt": "2026-10-04T13:50:35.090141+00:00" }, { "id": "arxiv-2210.05714", "title": "Visual Language Maps for Robot Navigation", "titleZh": "VLMaps：用于机器人导航的视觉语言地图", "date": "2022-10-11", "datePrecision": "day", "year": 2022, "url": "https://arxiv.org/abs/2210.05714", "paperUrl": "https://arxiv.org/abs/2210.05714", "codeUrl": "https://github.com/vlmaps/vlmaps", "summary": "把冻结LSeg的像素语言特征用RGB-D和相机位姿投影、跨视角融合成空间地图，文本嵌入检索开放词汇地标。语言模型生成调用预定义导航原语的代码，解析相对位置与重复动作；不同机器人可按类别生成不同障碍图。", "abstractZh": "把冻结LSeg的像素语言特征用RGB-D和相机位姿投影、跨视角融合成空间地图，文本嵌入检索开放词汇地标。语言模型生成调用预定义导航原语的代码，解析相对位置与重复动作；不同机器人可按类别生成不同障碍图。\nHabitat/Matterport3D十场景评测91组四目标序列，另七场景21条空间指令；目标容差一米。AI2THOR验证LoCoBot与无人机的不同通行图。真实平台是HSR，374帧建图后测试20个语言目标，使用RTAB-Map定位。", "category": "语义建图 / 开放词汇导航", "tags": [ "VLMaps", "视觉语言地图", "开放词汇", "空间关系" ], "directions": [ "导航与建图" ], "tier": "classic", "experimentType": "both", "robots": [ "Toyota HSR", "LoCoBot（仿真）", "无人机（仿真，型号未注明）" ], "robotFilters": [ "Toyota HSR", "LoCoBot" ], "codeStatus": "open", "status": "实现已开源", "trainingNote": "官方 MIT 代码提供建图与导航应用；使用预训练视觉语言特征，不要求为地图查询重新采集标签或微调模型。依赖模型/数据需分别授权。", "contribution": "把冻结LSeg的像素语言特征用RGB-D和相机位姿投影、跨视角融合成空间地图，文本嵌入检索开放词汇地标。语言模型生成调用预定义导航原语的代码，解析相对位置与重复动作；不同机器人可按类别生成不同障碍图。", "whyUseful": "需保存图像深度标定、地图栅格、特征融合和LLM原语提示，分别计连续链成功、独立子目标与SPL。不要把LoCoBot/无人机仿真写成实机；本次未部署导航。", "limitations": "依赖先建立环境地图、深度和位姿精度；相似物体歧义、重建噪声和动作漂移会失败，未处理持续动态场景。零样本主要指语言/语义组合，不是完全未知地图下自主探索。", "license": "MIT（VLMaps 源码；依赖另核）", "evidence": [ { "url": "https://arxiv.org/abs/2210.05714", "note": "首版日期、地图表示及仿真/实机范围。" }, { "url": "https://arxiv.org/html/2210.05714v2", "note": "§IV-E 明确 HSR、RTAB-Map、374 帧建图及 20 个目标中完成 10 个。" }, { "url": "https://github.com/vlmaps/vlmaps", "note": "作者官方实现及 MIT 许可证。" }, { "url": "https://arxiv.org/pdf/2210.05714", "note": "§III; Tables I–II：预建语义地图、导航原语、1米容差和链式成功。" }, { "url": "https://arxiv.org/pdf/2210.05714", "note": "§IV-E; §V：HSR实机10/20，噪声与动态物体限制。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "把语言语义嵌入可度量地图，让开放词汇与空间关系能够参与导航规划。", "experimentNote": "Habitat/Matterport3D十场景评测91组四目标序列，另七场景21条空间指令；目标容差一米。AI2THOR验证LoCoBot与无人机的不同通行图。真实平台是HSR，374帧建图后测试20个语言目标，使用RTAB-Map定位。", "projectUrl": null, "shortTitle": "", "original": { "id": "arxiv-2210.05714", "originalSourceUrl": "https://arxiv.org/pdf/2210.05714v4", "versionedOriginalUrl": "https://arxiv.org/pdf/2210.05714v4", "finalDownloadUrl": "https://arxiv.org/pdf/2210.05714", "sha256": "c7bb57c0defb3761526f740d187d215980b6a757eec14c7948d2d8c73c7b421d", "bytes": 6821714, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:01.464670+00:00", "publicationDate": "2022/10/11", "publicationDateSourceUrl": "https://arxiv.org/abs/2210.05714", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2210.05714", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2210.05714", "pages": 11, "sourceIdentityStatus": "title_match", "sourceTitle": "Visual Language Maps for Robot Navigation", "sourceVersion": "2210.05714v4", "sourceVersionDate": "2023/03/08", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.198818+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2210.05714" }, "originalAnalysis": { "id": "arxiv-2210.05714", "sourceUrl": "https://arxiv.org/pdf/2210.05714", "sectionsRead": [ "§III-A–D", "§IV-A–E", "Tables I–III", "§V" ], "methodsZh": "把冻结LSeg的像素语言特征用RGB-D和相机位姿投影、跨视角融合成空间地图，文本嵌入检索开放词汇地标。语言模型生成调用预定义导航原语的代码，解析相对位置与重复动作；不同机器人可按类别生成不同障碍图。", "experimentsZh": "Habitat/Matterport3D十场景评测91组四目标序列，另七场景21条空间指令；目标容差一米。AI2THOR验证LoCoBot与无人机的不同通行图。真实平台是HSR，374帧建图后测试20个语言目标，使用RTAB-Map定位。", "resultsZh": "作者报告多目标连续达成1至4个的成功率59%、34%、22%、15%，空间指令对应62%、33%、14%、10%；真实HSR完成10/20。跨形态障碍图改善部分路径效率，但四目标链成功仍低，不能概括为可靠长程导航。", "limitationsZh": "依赖先建立环境地图、深度和位姿精度；相似物体歧义、重建噪声和动作漂移会失败，未处理持续动态场景。零样本主要指语言/语义组合，不是完全未知地图下自主探索。", "reproductionZh": "需保存图像深度标定、地图栅格、特征融合和LLM原语提示，分别计连续链成功、独立子目标与SPL。不要把LoCoBot/无人机仿真写成实机；本次未部署导航。", "experimentType": "both", "robots": [ "Toyota HSR", "LoCoBot（仿真）", "无人机（仿真，型号未注明）" ], "evidenceNotes": [ { "section/page": "§III; Tables I–II", "note": "预建语义地图、导航原语、1米容差和链式成功。" }, { "section/page": "§IV-E; §V", "note": "HSR实机10/20，噪声与动态物体限制。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2210.03370", "title": "GNM: A General Navigation Model to Drive Any Robot", "date": "2022-10-07", "url": "https://arxiv.org/abs/2210.03370v2", "titleZh": "GNM：跨机器人的通用视觉导航模型", "abstractZh": "把不同机器人的动作统一为相对航点和偏航变化，用历史图像推断运动特性；联合预测目标距离与动作，并与拓扑地图组合。\n六种机器人约60小时轨迹训练，部署四种平台；Tello被限制在离地一米的水平面，Jackal和LoCoBot在20种室内外环境比较。", "summary": "把不同机器人的动作统一为相对航点和偏航变化，用历史图像推断运动特性；联合预测目标距离与动作，并与拓扑地图组合。", "experimentType": "real", "robots": [ "Clearpath Jackal", "LoCoBot（Kobuki底盘）", "Vizbot（Roomba底盘）", "DJI Tello" ], "tags": [ "导航与建图", "通用机器人策略", "模仿学习" ], "limitations": "未显式建模不同机体的全部通行能力，默认前向RGB与类似地面动作；不支持任意传感器、全空间飞行或操纵动作。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的跨机器人导航学习研究，关注共享动作空间及迁移假设。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2210.03370v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2210.03370v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/robodhruv/visualnav-transformer", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/robodhruv/visualnav-transformer", "category": "导航与建图 / 通用机器人策略", "year": 2022, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "训练、微调、ROS 数据处理及部署脚本已发布。 列出公开数据；部分原训练数据未发布，需联系作者。 作者提供 GNM/ViNT/NoMaD 检查点链接。 作者明确训练混合了公开和未公开数据，不能宣称从零完全复现全部原始数据；不同方法共仓不应重复计数。 本次未运行训练、复现实验或实机控制。", "directions": [ "导航与建图", "通用机器人策略", "模仿学习" ], "robotFilters": [ "Clearpath Jackal", "LoCoBot（Kobuki底盘）", "Vizbot（Roomba底盘）", "DJI Tello" ], "original": { "id": "arxiv-2210.03370", "originalSourceUrl": "https://arxiv.org/pdf/2210.03370v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2210.03370v2", "finalDownloadUrl": "https://arxiv.org/pdf/2210.03370v2", "sha256": "18b4f3c35feb35874d6db54d5f3f0eca9535c546877cfa747a3524fc5f841c3a", "bytes": 5181975, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:01:35.786862+00:00", "publicationDate": "2022/10/07", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2210.03370v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2210.03370v2", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "GNM: A General Navigation Model to Drive Any Robot", "sourceVersion": "2210.03370v2", "sourceVersionDate": "2023/05/22", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2210.03370v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2210.03370" }, "originalAnalysis": { "id": "arxiv-2210.03370", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2210.03370v2", "sourceVersion": "2210.03370v2", "sectionsRead": [ "PDF第3页选段（提取文本L308–325）", "PDF第3页选段（提取文本L336–349）", "PDF第3–4页选段（提取文本L383–400）", "PDF第4页选段（提取文本L448–461）", "PDF第6页选段（提取文本L797–814）", "PDF第4页选段（提取文本L410–446）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "把不同机器人的动作统一为相对航点和偏航变化，用历史图像推断运动特性；联合预测目标距离与动作，并与拓扑地图组合。", "experimentsZh": "六种机器人约60小时轨迹训练，部署四种平台；Tello被限制在离地一米的水平面，Jackal和LoCoBot在20种室内外环境比较。", "resultsZh": "多机器人训练优于单域策略，并迁移到未参与训练的平台；成功展示的是共享地面导航结构下的迁移。", "limitationsZh": "未显式建模不同机体的全部通行能力，默认前向RGB与类似地面动作；不支持任意传感器、全空间飞行或操纵动作。", "reproductionZh": "保留速度缩放、历史观测、正负图像对与图连通阈值；分开报告新机体和已见机体新环境的迁移结果。", "experimentType": "real", "robots": [ "Clearpath Jackal", "LoCoBot（Kobuki底盘）", "Vizbot（Roomba底盘）", "DJI Tello" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "把不同机器人的动作统一为相对航点和偏航变化，用历史图像推断运动特性；联合预测目标距离与动作，并与拓扑地图组合。" }, { "section": "原文实验与结果相关选段", "note": "六种机器人约60小时轨迹训练，部署四种平台；Tello被限制在离地一米的水平面，Jackal和LoCoBot在20种室内外环境比较。 多机器人训练优于单域策略，并迁移到未参与训练的平台；成功展示的是共享地面导航结构下的迁移。" }, { "section": "局限与复现条件", "note": "未显式建模不同机体的全部通行能力，默认前向RGB与类似地面动作；不支持任意传感器、全空间飞行或操纵动作。" } ], "originalSha256": "18b4f3c35feb35874d6db54d5f3f0eca9535c546877cfa747a3524fc5f841c3a", "readingScopeSha256": "ce2ed55d774887a4308b54b91ed49467a5802088257a5c7b2af3ea8e7a6cb9ab", "textExtractionSha256": "408a06bccb422011a684718fdd7a83424a6c1832ddd011e8bd6606a76f754915" }, "experimentNote": "六种机器人约60小时轨迹训练，部署四种平台；Tello被限制在离地一米的水平面，Jackal和LoCoBot在20种室内外环境比较。", "contribution": "把不同机器人的动作统一为相对航点和偏航变化，用历史图像推断运动特性；联合预测目标距离与动作，并与拓扑地图组合。", "whyUseful": "保留速度缩放、历史观测、正负图像对与图连通阈值；分开报告新机体和已见机体新环境的迁移结果。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "主代码 MIT；预训练权重和多来源数据仍需逐项查明条款。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2210.02697", "title": "DexGraspNet: A Large-Scale Robotic Dexterous Grasp Dataset for General Objects Based on Simulation", "date": "2022-10-06", "url": "https://arxiv.org/abs/2210.02697v2", "titleZh": "DexGraspNet：大规模仿真生成的灵巧抓取数据集", "abstractZh": "用可微力闭合能量、手姿初始化和几何穿透约束加速优化抓取；再在Isaac Gym中验证不同重力方向下的稳定性。\n为5355个物体生成132万条Shadow Hand抓取，比较在DexGraspNet与旧数据集上训练的抓取网络，并报告质量、多样性和穿透。", "summary": "用可微力闭合能量、手姿初始化和几何穿透约束加速优化抓取；再在Isaac Gym中验证不同重力方向下的稳定性。", "experimentType": "sim", "robots": [ "Shadow Hand（仿真）", "Allegro Hand（生成示例）" ], "tags": [ "数据集与基准", "灵巧手", "操作与抓取" ], "limitations": "优化倾向多接触的包络抓取，精细抓取稀少；无语义约束，不能保证按杯柄等功能部位抓取，也未构成实机验证。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的灵巧抓取数据生成研究，重点阅读力闭合约束与数据评测协议。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "源码可用，存在使用限制", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2210.02697v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2210.02697v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/PKU-EPIC/DexGraspNet", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "source_available", "codeUrl": "https://github.com/PKU-EPIC/DexGraspNet", "category": "数据集与基准 / 灵巧手", "year": 2022, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "提供数据生成/仿真验证，而非完整通用机器人策略训练。 作者项目页提供完整数据下载链接。 不以预训练机器人策略权重为主要发布物。 抓取数据和生成流程不等于可直接部署的视觉抓取策略；物体缩放、碰撞和控制需另核验。 本次未运行训练、复现实验或实机控制。", "directions": [ "数据集与基准", "灵巧手", "操作与抓取" ], "robotFilters": [ "Shadow Hand（仿真）", "Allegro Hand（生成示例）" ], "original": { "id": "arxiv-2210.02697", "originalSourceUrl": "https://arxiv.org/pdf/2210.02697v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2210.02697v2", "finalDownloadUrl": "https://arxiv.org/pdf/2210.02697v2", "sha256": "c1b5fe6258e25369626a6ac1df06218394bbb3fb3aa4bb41980dca074e6c04b8", "bytes": 4584838, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:14.649001+00:00", "publicationDate": "2022/10/06", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2210.02697v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2210.02697v2", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "DexGraspNet: A Large-Scale Robotic Dexterous Grasp Dataset for General Objects Based on Simulation", "sourceVersion": "2210.02697v2", "sourceVersionDate": "2023/03/08", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2210.02697v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2210.02697" }, "originalAnalysis": { "id": "arxiv-2210.02697", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2210.02697v2", "sourceVersion": "2210.02697v2", "sectionsRead": [ "PDF第3页选段（提取文本L364–381）", "PDF第4页选段（提取文本L478–491）", "PDF第6页选段（提取文本L750–767）", "PDF第3页选段（提取文本L264–309）", "PDF第6页选段（提取文本L661–701）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "用可微力闭合能量、手姿初始化和几何穿透约束加速优化抓取；再在Isaac Gym中验证不同重力方向下的稳定性。", "experimentsZh": "为5355个物体生成132万条Shadow Hand抓取，比较在DexGraspNet与旧数据集上训练的抓取网络，并报告质量、多样性和穿透。", "resultsZh": "新数据训练改善文中基线；注意数据过滤要求六方向均稳定，而基线评估放宽为至少一个方向稳定，两种成功定义不同。", "limitationsZh": "优化倾向多接触的包络抓取，精细抓取稀少；无语义约束，不能保证按杯柄等功能部位抓取，也未构成实机验证。", "reproductionZh": "须固定物体尺度、碰撞网格、摩擦和成功判据，区分生成数据过滤与学习方法评估的重力方向及穿透阈值。", "experimentType": "sim", "robots": [ "Shadow Hand（仿真）", "Allegro Hand（生成示例）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "用可微力闭合能量、手姿初始化和几何穿透约束加速优化抓取；再在Isaac Gym中验证不同重力方向下的稳定性。" }, { "section": "原文实验与结果相关选段", "note": "为5355个物体生成132万条Shadow Hand抓取，比较在DexGraspNet与旧数据集上训练的抓取网络，并报告质量、多样性和穿透。 新数据训练改善文中基线；注意数据过滤要求六方向均稳定，而基线评估放宽为至少一个方向稳定，两种成功定义不同。" }, { "section": "局限与复现条件", "note": "优化倾向多接触的包络抓取，精细抓取稀少；无语义约束，不能保证按杯柄等功能部位抓取，也未构成实机验证。" } ], "originalSha256": "c1b5fe6258e25369626a6ac1df06218394bbb3fb3aa4bb41980dca074e6c04b8", "readingScopeSha256": "161743a063b80cedf89717f01d70af597179c4c22fa7886082063a30c22175c1", "textExtractionSha256": "4b05af86f382c7db0f42e41ff6c8ea236d99aea2b9b4ce40c36cfbeac8c3154a" }, "experimentNote": "为5355个物体生成132万条Shadow Hand抓取，比较在DexGraspNet与旧数据集上训练的抓取网络，并报告质量、多样性和穿透。", "contribution": "用可微力闭合能量、手姿初始化和几何穿透约束加速优化抓取；再在Isaac Gym中验证不同重力方向下的稳定性。", "whyUseful": "须固定物体尺度、碰撞网格、摩擦和成功判据，区分生成数据过滤与学习方法评估的重力方向及穿透阈值。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "README 明确代码工作与数据集均为 CC BY-NC 4.0，禁止商业用途，不能标成 MIT 等宽松开源。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2210.03094", "title": "VIMA: General Robot Manipulation with Multimodal Prompts", "titleZh": "VIMA：利用多模态提示完成通用机器人操作", "shortTitle": "VIMA", "date": "2022-10-06", "year": 2022, "url": "https://arxiv.org/abs/2210.03094", "codeUrl": "https://github.com/vimalabs/VIMA", "category": "视觉语言动作", "tags": [ "多模态提示", "对象中心", "模仿学习", "仿真" ], "tier": "foundation", "summary": "VIMA把任务写成文本与对象/场景图像交错的多模态提示，冻结T5编码提示，通过交叉注意力条件化自回归动作解码。观察采用对象检测得到的图像块和位置词元，并保留动作—观察历史，统一表达新概念、视觉目标和视频模仿。", "abstractZh": "VIMA把任务写成文本与对象/场景图像交错的多模态提示，冻结T5编码提示，通过交叉注意力条件化自回归动作解码。观察采用对象检测得到的图像块和位置词元，并保留动作—观察历史，统一表达新概念、视觉目标和视频模仿。\nVIMA-Bench是PyBullet UR5桌面仿真，17类任务、60万余脚本专家轨迹，动作是带两个SE(2)目标的抓放或推擦原语。四级评估依次改变摆放、组合、对象和任务，比较改造后的Gato/Flamingo/GPT架构及模型/数据规模。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "sim", "robots": [ "Universal Robots UR5（仿真）" ], "codeStatus": "open", "status": "MIT 代码、预训练模型、数据与 VIMA-Bench 已公开。", "trainingNote": "官方仓库提供模型与评测脚本；数据约 60 万条专家轨迹来自仿真。", "whyUseful": "固定四级留出划分、oracle原语与检测器训练数据，分别测检测错误、对象词元和提示交叉注意力；不要将oracle分割与预测分割混用。比较总参数与预训练成本时补上T5及视觉管线，独立报告各级任务成功。", "contribution": "VIMA把任务写成文本与对象/场景图像交错的多模态提示，冻结T5编码提示，通过交叉注意力条件化自回归动作解码。观察采用对象检测得到的图像块和位置词元，并保留动作—观察历史，统一表达新概念、视觉目标和视频模仿。", "limitations": "没有本篇实机验证；高层动作原语掩盖低层接触控制难度。依赖独立检测器，遮挡和未见形状会传播误差；简化桌面物理与模板提示不等于开放世界通用机器人语言理解。", "caveats": "预印本标题含 General；ICML 发表题名为 VIMA: Robot Manipulation with Multimodal Prompts。", "evidence": [ { "url": "https://arxiv.org/abs/2210.03094", "note": "初版日期与摘要。" }, { "url": "https://arxiv.org/html/2210.03094v2", "note": "附录描述 PyBullet、Ravens 与 UR5 仿真。" }, { "url": "https://github.com/vimalabs/VIMA", "note": "官方模型实现、MIT 许可证、权重和数据入口。" }, { "url": "https://arxiv.org/pdf/2210.03094", "note": "3–4; Appendix A：17任务、原语接口和PyBullet UR5。" }, { "url": "https://arxiv.org/pdf/2210.03094", "note": "5; Appendix G：架构/规模对比、T5计数及检测器/仿真限制。" }, { "url": "https://arxiv.org/pdf/2210.03094v2", "note": "3–4; Appendix A：17任务、原语接口和PyBullet UR5。" }, { "url": "https://arxiv.org/pdf/2210.03094v2", "note": "5; Appendix G：架构/规模对比、T5计数及检测器/仿真限制。" } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "操作与抓取", "模仿学习" ], "robotFilters": [ "Universal Robots UR5" ], "verificationNote": "已核论文、仿真平台和官方公开代码；无该论文实机评测证据。", "original": { "id": "arxiv-2210.03094", "originalSourceUrl": "https://arxiv.org/pdf/2210.03094v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2210.03094v2", "finalDownloadUrl": "https://arxiv.org/pdf/2210.03094", "sha256": "cf849f80368749b0310b43117519ce0abff3e8ae49744e78c4a29b9d17f0695f", "bytes": 30270724, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:05.060828+00:00", "publicationDate": "2022/10/06", "publicationDateSourceUrl": "https://arxiv.org/abs/2210.03094", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2210.03094", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2210.03094", "pages": 48, "sourceIdentityStatus": "title_match", "sourceTitle": "VIMA: General Robot Manipulation with Multimodal Prompts", "sourceVersion": "2210.03094v2", "sourceVersionDate": "2023/05/28", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.227766+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2210.03094" }, "originalAnalysis": { "id": "arxiv-2210.03094", "sourceUrl": "https://arxiv.org/pdf/2210.03094v2", "sectionsRead": [ "2–4", "5.1–5.3", "Appendix A,G", "Fig.4" ], "methodsZh": "VIMA把任务写成文本与对象/场景图像交错的多模态提示，冻结T5编码提示，通过交叉注意力条件化自回归动作解码。观察采用对象检测得到的图像块和位置词元，并保留动作—观察历史，统一表达新概念、视觉目标和视频模仿。", "experimentsZh": "VIMA-Bench是PyBullet UR5桌面仿真，17类任务、60万余脚本专家轨迹，动作是带两个SE(2)目标的抓放或推擦原语。四级评估依次改变摆放、组合、对象和任务，比较改造后的Gato/Flamingo/GPT架构及模型/数据规模。", "resultsZh": "对象词元与交叉注意力的结合在测试容量范围最稳健，少量数据时优势尤其大；最难泛化设定作者报告最高约2.9倍成功，1%数据在部分设定可匹配基线10%数据。模型规模曲线只计控制器，固定111M的T5编码器未计入。", "limitationsZh": "没有本篇实机验证；高层动作原语掩盖低层接触控制难度。依赖独立检测器，遮挡和未见形状会传播误差；简化桌面物理与模板提示不等于开放世界通用机器人语言理解。", "reproductionZh": "固定四级留出划分、oracle原语与检测器训练数据，分别测检测错误、对象词元和提示交叉注意力；不要将oracle分割与预测分割混用。比较总参数与预训练成本时补上T5及视觉管线，独立报告各级任务成功。", "experimentType": "sim", "robots": [ "Universal Robots UR5（仿真）" ], "corrections": [ "只验证仿真，高层抓放/推擦原语不是逐关节视觉闭环控制；控制器参数未计T5编码器。" ], "evidenceNotes": [ { "note": "17任务、原语接口和PyBullet UR5。", "section": "3–4; Appendix A" }, { "note": "架构/规模对比、T5计数及检测器/仿真限制。", "section": "5; Appendix G" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2210.03094v2", "originalSha256": "cf849f80368749b0310b43117519ce0abff3e8ae49744e78c4a29b9d17f0695f" }, "experimentNote": "VIMA-Bench是PyBullet UR5桌面仿真，17类任务、60万余脚本专家轨迹，动作是带两个SE(2)目标的抓放或推擦原语。四级评估依次改变摆放、组合、对象和任务，比较改造后的Gato/Flamingo/GPT架构及模型/数据规模。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2209.07753", "title": "Code as Policies: Language Model Programs for Embodied Control", "date": "2022-09-16", "url": "https://arxiv.org/abs/2209.07753v4", "titleZh": "代码即策略：面向具身控制的语言模型程序", "abstractZh": "用少样本提示将自然语言变成Python策略程序，组合感知API和控制原语；遇到未定义函数时递归生成子程序。\n包含37题机器人代码生成基准、HumanEval、仿真桌面控制，以及UR5e绘图和真实桌面抓放；代码测试与物理成功需分开解读。", "summary": "用少样本提示将自然语言变成Python策略程序，组合感知API和控制原语；遇到未定义函数时递归生成子程序。", "experimentType": "both", "robots": [ "Universal Robots UR5e" ], "tags": [ "视觉语言动作", "操作与抓取" ], "limitations": "能力受感知API和控制原语覆盖限制；长且复杂的指令、跨抽象层任务较脆弱，无法事先保证生成响应正确。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的语言模型生成机器人程序研究，用于理解分层API控制的能力和风险。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2209.07753v4", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2209.07753v4", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/google-research/google-research/tree/master/code_as_policies", "note": "已发布 notebook 与实验示例；非完整实机控制栈；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/google-research/google-research/tree/master/code_as_policies", "category": "视觉语言动作 / 操作与抓取", "year": 2022, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "发布提示/代码生成及实验 notebook；不含基础 LLM 预训练。 提供示例和基准 notebook，不是机器人示范数据集发布。 没有自有基础模型权重发布。 代码示例不等于公开了所用语言模型的训练数据或权重；实机控制 API 仍需用户实现并设置安全边界。 本次未运行训练、复现实验或实机控制。", "directions": [ "视觉语言动作", "操作与抓取" ], "robotFilters": [ "Universal Robots UR5e" ], "original": { "id": "arxiv-2209.07753", "originalSourceUrl": "https://arxiv.org/pdf/2209.07753v4", "versionedOriginalUrl": "https://arxiv.org/pdf/2209.07753v4", "finalDownloadUrl": "https://arxiv.org/pdf/2209.07753v4", "sha256": "e36857fa13746da031444ec59549e61204ca92d599b1d048ac1c55aaf1f6d0a3", "bytes": 3200561, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T14:59:50.163040+00:00", "publicationDate": "2022/09/16", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2209.07753v4", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2209.07753v4", "pages": 16, "sourceIdentityStatus": "title_match", "sourceTitle": "Code as Policies: Language Model Programs for Embodied Control", "sourceVersion": "2209.07753v4", "sourceVersionDate": "2023/05/25", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2209.07753v4", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2209.07753" }, "originalAnalysis": { "id": "arxiv-2209.07753", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2209.07753v4", "sourceVersion": "2209.07753v4", "sectionsRead": [ "PDF第2页选段（提取文本L195–212）", "PDF第5页选段（提取文本L617–634）", "PDF第10页选段（提取文本L1103–1120）", "PDF第12页选段（提取文本L1386–1399）", "PDF第3页选段（提取文本L269–302）", "PDF第6–7页选段（提取文本L776–812）", "PDF第13页选段（提取文本L1472–1498）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "用少样本提示将自然语言变成Python策略程序，组合感知API和控制原语；遇到未定义函数时递归生成子程序。", "experimentsZh": "包含37题机器人代码生成基准、HumanEval、仿真桌面控制，以及UR5e绘图和真实桌面抓放；代码测试与物理成功需分开解读。", "resultsZh": "分层生成相较平铺生成提高文中代码测试表现，并展示空间推理与开放词汇指令组合。", "limitationsZh": "能力受感知API和控制原语覆盖限制；长且复杂的指令、跨抽象层任务较脆弱，无法事先保证生成响应正确。", "reproductionZh": "保留提示词、API语义和模型版本；先在隔离环境做静态检查与单元测试，再验证动作边界，禁止直接将任意代码送实机执行。", "experimentType": "both", "robots": [ "Universal Robots UR5e" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "用少样本提示将自然语言变成Python策略程序，组合感知API和控制原语；遇到未定义函数时递归生成子程序。" }, { "section": "原文实验与结果相关选段", "note": "包含37题机器人代码生成基准、HumanEval、仿真桌面控制，以及UR5e绘图和真实桌面抓放；代码测试与物理成功需分开解读。 分层生成相较平铺生成提高文中代码测试表现，并展示空间推理与开放词汇指令组合。" }, { "section": "局限与复现条件", "note": "能力受感知API和控制原语覆盖限制；长且复杂的指令、跨抽象层任务较脆弱，无法事先保证生成响应正确。" } ], "originalSha256": "e36857fa13746da031444ec59549e61204ca92d599b1d048ac1c55aaf1f6d0a3", "readingScopeSha256": "c2a658876bdea7f15aefb9d366af95cc719adf68faf135f553d858f070143030", "textExtractionSha256": "d6c360c4a71a05844b290b63fe1349489830c7291de3fd0bb6fc4f90feb4e9a2" }, "experimentNote": "包含37题机器人代码生成基准、HumanEval、仿真桌面控制，以及UR5e绘图和真实桌面抓放；代码测试与物理成功需分开解读。", "contribution": "用少样本提示将自然语言变成Python策略程序，组合感知API和控制原语；遇到未定义函数时递归生成子程序。", "whyUseful": "保留提示词、API语义和模型版本；先在隔离环境做静态检查与单元测试，再验证动作边界，禁止直接将任意代码送实机执行。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "Google Research 仓库代码 Apache-2.0；模型服务和第三方依赖不随代码许可开放。", "codeReleaseScope": "已发布 notebook 与实验示例；非完整实机控制栈" }, { "id": "arxiv-2209.05451", "title": "Perceiver-Actor: A Multi-Task Transformer for Robotic Manipulation", "date": "2022-09-12", "url": "https://arxiv.org/abs/2209.05451v2", "titleZh": "PerAct：用于多任务机器人操作的感知器动作Transformer", "abstractZh": "将多视角RGB-D体素化，用Perceiver潜变量融合语言与三维场景，预测离散末端位置、旋转和夹爪状态，由运动规划器执行关键帧动作。\n仿真使用18个RLBench任务、249个变体，比较每任务10或100条示教；每任务25个评估回合，另有真实机器人验证。", "summary": "将多视角RGB-D体素化，用Perceiver潜变量融合语言与三维场景，预测离散末端位置、旋转和夹爪状态，由运动规划器执行关键帧动作。", "experimentType": "both", "robots": [ "Franka Panda" ], "tags": [ "操作与抓取", "模仿学习", "视觉语言动作" ], "limitations": "依赖相机标定、体素离散精度和采样式运动规划；连续灵巧控制、多指高自由度执行器不直接适用。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的三维语言条件操作方法，用于比较体素动作表示与后续多视图策略。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2209.05451v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2209.05451v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/peract/peract", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/peract/peract", "category": "操作与抓取 / 模仿学习", "year": 2022, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "train.py 与 eval.py 已发布，含多 GPU 配置。 预生成 RLBench train/val/test 与真实机器人数据链接已列出。 官方发布 peract_600k.zip 最终检查点。 公开最终检查点不一定是每任务的最优检查点；仿真依赖指定分支，真实碰撞规划需另接 MoveIt。 本次未运行训练、复现实验或实机控制。", "directions": [ "操作与抓取", "模仿学习", "视觉语言动作" ], "robotFilters": [ "Franka Panda" ], "original": { "id": "arxiv-2209.05451", "originalSourceUrl": "https://arxiv.org/pdf/2209.05451v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2209.05451v2", "finalDownloadUrl": "https://arxiv.org/pdf/2209.05451v2", "sha256": "daab02c1ba8c47505f9eb54e8e61fd327e799bb20b392efd7f4d06b0357d0f17", "bytes": 8181953, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T14:59:39.854116+00:00", "publicationDate": "2022/09/12", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2209.05451v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2209.05451v2", "pages": 28, "sourceIdentityStatus": "title_match", "sourceTitle": "Perceiver-Actor: A Multi-Task Transformer for Robotic Manipulation", "sourceVersion": "2209.05451v2", "sourceVersionDate": "2022/11/11", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2209.05451v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2209.05451" }, "originalAnalysis": { "id": "arxiv-2209.05451", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2209.05451v2", "sourceVersion": "2209.05451v2", "sectionsRead": [ "PDF第6页选段（提取文本L404–421）", "PDF第6页选段（提取文本L427–440）", "PDF第8页选段（提取文本L993–1010）", "PDF第3–4页选段（提取文本L190–243）", "PDF第7页选段（提取文本L871–891）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "将多视角RGB-D体素化，用Perceiver潜变量融合语言与三维场景，预测离散末端位置、旋转和夹爪状态，由运动规划器执行关键帧动作。", "experimentsZh": "仿真使用18个RLBench任务、249个变体，比较每任务10或100条示教；每任务25个评估回合，另有真实机器人验证。", "resultsZh": "体素动作表示相比直接图像到动作基线更具数据效率；任务完成按完整成功计分，没有部分成功分数。", "limitationsZh": "依赖相机标定、体素离散精度和采样式运动规划；连续灵巧控制、多指高自由度执行器不直接适用。", "reproductionZh": "须保持关键帧抽取、语言模板、数据划分和规划器一致；固定体素分辨率，不能混用任务单独选优与统一多任务模型。", "experimentType": "both", "robots": [ "Franka Panda" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "将多视角RGB-D体素化，用Perceiver潜变量融合语言与三维场景，预测离散末端位置、旋转和夹爪状态，由运动规划器执行关键帧动作。" }, { "section": "原文实验与结果相关选段", "note": "仿真使用18个RLBench任务、249个变体，比较每任务10或100条示教；每任务25个评估回合，另有真实机器人验证。 体素动作表示相比直接图像到动作基线更具数据效率；任务完成按完整成功计分，没有部分成功分数。" }, { "section": "局限与复现条件", "note": "依赖相机标定、体素离散精度和采样式运动规划；连续灵巧控制、多指高自由度执行器不直接适用。" } ], "originalSha256": "daab02c1ba8c47505f9eb54e8e61fd327e799bb20b392efd7f4d06b0357d0f17", "readingScopeSha256": "ea45714d4ae940d7167c94b5d774c22d774443f79aa3d19283e54f57a9cde73d", "textExtractionSha256": "62482c20d4f2d5c6ac7153540b13b0a2eb45ce7db9e80ce0fbf0004e282ac55c" }, "experimentNote": "仿真使用18个RLBench任务、249个变体，比较每任务10或100条示教；每任务25个评估回合，另有真实机器人验证。", "contribution": "将多视角RGB-D体素化，用Perceiver潜变量融合语言与三维场景，预测离散末端位置、旋转和夹爪状态，由运动规划器执行关键帧动作。", "whyUseful": "须保持关键帧抽取、语言模板、数据划分和规划器一致；固定体素分辨率，不能混用任务单独选优与统一多任务模型。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "PerAct 核心 Apache-2.0；ARM、RLBench、CoppeliaSim 等依赖各有条款，不能视整套为单一 Apache 许可。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2209.00588", "title": "Transformers are Sample-Efficient World Models", "titleZh": "Transformer 是样本高效的世界模型", "shortTitle": "IRIS", "date": "2022-09-01", "year": 2022, "url": "https://arxiv.org/abs/2209.00588", "paperUrl": "https://arxiv.org/abs/2209.00588", "projectUrl": null, "codeUrl": "https://github.com/eloialonso/iris", "summary": "IRIS以离散自编码器把64×64像素图像压缩成令牌，GPT式Transformer在图像令牌与动作交错序列上预测未来画面、奖励和终止。策略和值网络只在生成轨迹中用带熵奖励的actor-critic更新；环境交…", "abstractZh": "IRIS以离散自编码器把64×64像素图像压缩成令牌，GPT式Transformer在图像令牌与动作交错序列上预测未来画面、奖励和终止。策略和值网络只在生成轨迹中用带熵奖励的actor-critic更新；环境交互用于改进世界模型，部署不进行树搜索。归档v2的图像重建使用L1而非初版所写L2。\n在26个Atari 100k游戏中，每游戏允许十万次环境动作，约两小时游戏经验。作者各做五次训练，每次结束评估100局，并用分层自助法报告均值、中位数、四分位均值等置信区间；区分带前瞻搜索的MuZero/EfficientZero与纯学习基线。", "category": "世界模型 / 自回归建模", "tags": [ "IRIS", "世界模型", "强化学习" ], "directions": [ "世界模型", "强化学习" ], "tier": "classic", "experimentType": "sim", "experimentNote": "在26个Atari 100k游戏中，每游戏允许十万次环境动作，约两小时游戏经验。作者各做五次训练，每次结束评估100局，并用分层自助法报告均值、中位数、四分位均值等置信区间；区分带前瞻搜索的MuZero/EfficientZero与纯学习基线。", "robots": [], "robotFilters": [], "robotNote": "无实机；仿真形态不等同于已验证的商业机器人型号。", "codeStatus": "open", "status": "训练代码已开源", "trainingStatus": "训练代码已开源", "trainingNote": "官方训练代码、结果和预训练模型入口公开；代码 GPL-3.0，Atari ROM 另有许可要求。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "GPL-3.0（代码；ROM 另行许可）", "contribution": "IRIS以离散自编码器把64×64像素图像压缩成令牌，GPT式Transformer在图像令牌与动作交错序列上预测未来画面、奖励和终止。策略和值网络只在生成轨迹中用带熵奖励的actor-critic更新；环境交互用于改进世界模型，部署不进行树搜索。归档v2的图像重建使用L1而非初版所写L2。", "whyUseful": "应固定Atari动作重复、十万动作预算、五种子与最终100局协议，复现自编码器、十层Transformer及20步想象。v2附录报告两个环境共享一张A100训练约七天；两小时只是交互经验。奖励预测依任务用均方或交叉熵，不能机械照搬初版描述。", "limitations": "罕见事件必须先被环境探索发现，再被模型学会、策略重新探索，存在双重探索瓶颈。16令牌会漏掉迷宫中小目标；64令牌改善重建却增加序列长度。模型偏差直接限制策略，论文没有机器人或连续动作实机实验。", "caveats": "Atari 结论不能直接推及接触动力学；自回归预测误差和推理成本仍需考虑。", "evidence": [ { "url": "https://arxiv.org/abs/2209.00588", "note": "日期、方法和 Atari 100k 评估。" }, { "url": "https://github.com/eloialonso/iris", "note": "官方训练、结果、模型入口和 GPL-3.0；ROM 许可提示。" }, { "url": "https://arxiv.org/pdf/2209.00588v2", "note": "v2 §2.1–2.3：Discrete autoencoder uses L1 reconstruction; reward loss is MSE or cross-entropy depending on reward; actor-critic trains in imagination." }, { "url": "https://arxiv.org/pdf/2209.00588v2", "note": "v2 §3.2/Table1/PDF p6：Visually verified:26games,100k actions,5runs,100 final episodes;mean1.046,median0.289,IQM0.501." }, { "url": "https://arxiv.org/pdf/2209.00588v2", "note": "v2 §3.2/AppendixE：Double exploration;64tokens improve reconstruction and returns for Alien/Asterix/BankHeist,with higher compute." }, { "url": "https://arxiv.org/pdf/2209.00588v2", "note": "v2 AppendixG：Two Atari environments per A100 40GB take about7days;authors describe average3.5days/environment. Two hours denotes game interaction,not compute." } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2022：展示离散词元 Transformer 可承担高交互效率的世界模型学习。", "freshness": "经典工作", "original": { "id": "arxiv-2209.00588", "originalSourceUrl": "https://arxiv.org/pdf/2209.00588v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2209.00588v2", "finalDownloadUrl": "https://arxiv.org/pdf/2209.00588", "sha256": "6b583542ba35dc99a36127dd3b7650aac5d28479f3fe0e5e47f741768ae587b4", "bytes": 655408, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:04.422147+00:00", "publicationDate": "2022/09/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2209.00588", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2209.00588", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2209.00588", "pages": 21, "sourceIdentityStatus": "title_match", "sourceTitle": "Transformers are Sample-Efficient World Models", "sourceVersion": "2209.00588v2", "sourceVersionDate": "2023/03/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.231129+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2209.00588" }, "originalAnalysis": { "id": "arxiv-2209.00588", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739396+00:00", "sourceUrl": "https://arxiv.org/pdf/2209.00588v2", "sectionsRead": [ "归档v2第2.1–2.3节：离散化、动力学与想象学习", "归档v2第3.1–3.3节及表1：协议、结果与错误分析", "归档v2附录B/D/E/G：策略目标、循环、令牌消融和计算资源", "归档v2 PDF第6页表1可视复核" ], "methodsZh": "IRIS以离散自编码器把64×64像素图像压缩成令牌，GPT式Transformer在图像令牌与动作交错序列上预测未来画面、奖励和终止。策略和值网络只在生成轨迹中用带熵奖励的actor-critic更新；环境交互用于改进世界模型，部署不进行树搜索。归档v2的图像重建使用L1而非初版所写L2。", "experimentsZh": "在26个Atari 100k游戏中，每游戏允许十万次环境动作，约两小时游戏经验。作者各做五次训练，每次结束评估100局，并用分层自助法报告均值、中位数、四分位均值等置信区间；区分带前瞻搜索的MuZero/EfficientZero与纯学习基线。", "resultsZh": "作者表1报告人类归一化均值1.046、四分位均值0.501，26项中10项超过人类；中位数仅0.289。其总体优势主要针对不使用搜索的方法，EfficientZero平均与中位成绩更高，不能称所有游戏全面领先。Pong生成示例捕捉球轨迹及计分。", "limitationsZh": "罕见事件必须先被环境探索发现，再被模型学会、策略重新探索，存在双重探索瓶颈。16令牌会漏掉迷宫中小目标；64令牌改善重建却增加序列长度。模型偏差直接限制策略，论文没有机器人或连续动作实机实验。", "reproductionZh": "应固定Atari动作重复、十万动作预算、五种子与最终100局协议，复现自编码器、十层Transformer及20步想象。v2附录报告两个环境共享一张A100训练约七天；两小时只是交互经验。奖励预测依任务用均方或交叉熵，不能机械照搬初版描述。", "evidenceNotes": [ { "section/page": "v2 §2.1–2.3", "note": "Discrete autoencoder uses L1 reconstruction; reward loss is MSE or cross-entropy depending on reward; actor-critic trains in imagination." }, { "section/page": "v2 §3.2/Table1/PDF p6", "note": "Visually verified:26games,100k actions,5runs,100 final episodes;mean1.046,median0.289,IQM0.501." }, { "section/page": "v2 §3.2/AppendixE", "note": "Double exploration;64tokens improve reconstruction and returns for Alien/Asterix/BankHeist,with higher compute." }, { "section/page": "v2 AppendixG", "note": "Two Atari environments per A100 40GB take about7days;authors describe average3.5days/environment. Two hours denotes game interaction,not compute." } ], "experimentType": "sim", "robots": [], "corrections": [ "“两小时”是游戏环境交互量，不是训练用时；无实体机器人评测。", "全文版本已校对至v2；重建损失与奖励损失描述以归档v2为准。" ], "sourceVersion": "2209.00588v2", "originalSha256": "6b583542ba35dc99a36127dd3b7650aac5d28479f3fe0e5e47f741768ae587b4" }, "analysisVerifiedAt": "2026-10-04T13:47:58.739396+00:00" }, { "id": "arxiv-2207.05608", "title": "Inner Monologue: Embodied Reasoning through Planning with Language Models", "date": "2022-07-12", "url": "https://arxiv.org/abs/2207.05608v1", "titleZh": "内心独白：通过语言反馈实现具身规划推理", "abstractZh": "把成功检测、物体列表、场景描述或人类反馈持续加入LLM提示，使其从已有短时技能库中重新选取动作，形成闭环计划。\n包含Ravens仿真、UR5e真实桌面重排及真实移动操作，比较仅物体反馈、成功反馈和更完整场景反馈。", "summary": "把成功检测、物体列表、场景描述或人类反馈持续加入LLM提示，使其从已有短时技能库中重新选取动作，形成闭环计划。", "experimentType": "both", "robots": [ "Universal Robots UR5e", "Everyday Robots移动操作平台（未细分型号）" ], "tags": [ "视觉语言动作", "操作与抓取" ], "limitations": "部分实验使用脚本或人类作为理想场景描述器；错误成功检测及低层技能不足会限制上层推理，LLM也可能忽略反馈。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的反馈闭环语言规划研究；重点是反馈与技能覆盖假设。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2207.05608v1", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2207.05608v1", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "视觉语言动作 / 操作与抓取", "year": 2022, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "已核验原文；本轮未独立核实官方实现、训练入口及发布范围，不能据论文声明认定代码已开放。 本次未运行训练、复现实验或实机控制。", "directions": [ "视觉语言动作", "操作与抓取" ], "robotFilters": [ "Universal Robots UR5e", "Everyday Robots移动操作平台（未细分型号）" ], "original": { "id": "arxiv-2207.05608", "originalSourceUrl": "https://arxiv.org/pdf/2207.05608v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2207.05608v1", "finalDownloadUrl": "https://arxiv.org/pdf/2207.05608v1", "sha256": "ecf30268add2af672291ad65b53c9fd87801f6490b0af195b12db1dedee0f2b9", "bytes": 10805774, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T14:59:50.319562+00:00", "publicationDate": "2022/07/12", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2207.05608v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2207.05608v1", "pages": 25, "sourceIdentityStatus": "title_match", "sourceTitle": "Inner Monologue: Embodied Reasoning through Planning with Language Models", "sourceVersion": "2207.05608v1", "sourceVersionDate": "2022/07/12", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2207.05608v1", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2207.05608" }, "originalAnalysis": { "id": "arxiv-2207.05608", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2207.05608v1", "sourceVersion": "2207.05608v1", "sectionsRead": [ "PDF第3页选段（提取文本L169–186）", "PDF第5页选段（提取文本L398–415）", "PDF第6页选段（提取文本L499–512）", "PDF第6页选段（提取文本L513–526）", "PDF第9页选段（提取文本L864–881）", "PDF第7页选段（提取文本L582–599）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "把成功检测、物体列表、场景描述或人类反馈持续加入LLM提示，使其从已有短时技能库中重新选取动作，形成闭环计划。", "experimentsZh": "包含Ravens仿真、UR5e真实桌面重排及真实移动操作，比较仅物体反馈、成功反馈和更完整场景反馈。", "resultsZh": "场景反馈支持故障后重试与重新规划，未见长指令表现优于文中无分层基线；结论依赖反馈来源。", "limitationsZh": "部分实验使用脚本或人类作为理想场景描述器；错误成功检测及低层技能不足会限制上层推理，LLM也可能忽略反馈。", "reproductionZh": "应标注每种反馈是人工、真值还是学习模块，并保留技能库覆盖范围和重试预算；不能将含人工反馈系统称为全自主。", "experimentType": "both", "robots": [ "Universal Robots UR5e", "Everyday Robots移动操作平台（未细分型号）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "把成功检测、物体列表、场景描述或人类反馈持续加入LLM提示，使其从已有短时技能库中重新选取动作，形成闭环计划。" }, { "section": "原文实验与结果相关选段", "note": "包含Ravens仿真、UR5e真实桌面重排及真实移动操作，比较仅物体反馈、成功反馈和更完整场景反馈。 场景反馈支持故障后重试与重新规划，未见长指令表现优于文中无分层基线；结论依赖反馈来源。" }, { "section": "局限与复现条件", "note": "部分实验使用脚本或人类作为理想场景描述器；错误成功检测及低层技能不足会限制上层推理，LLM也可能忽略反馈。" } ], "originalSha256": "ecf30268add2af672291ad65b53c9fd87801f6490b0af195b12db1dedee0f2b9", "readingScopeSha256": "05b8377168a3e423c5fb9aa84a941ad0cc58a6a51d2797645d103b2e09bd766a", "textExtractionSha256": "1325c7846bcd8bf2e9068406d773522717e0517cc13edc7ed6e9da5471e21be7" }, "experimentNote": "包含Ravens仿真、UR5e真实桌面重排及真实移动操作，比较仅物体反馈、成功反馈和更完整场景反馈。", "contribution": "把成功检测、物体列表、场景描述或人类反馈持续加入LLM提示，使其从已有短时技能库中重新选取动作，形成闭环计划。", "whyUseful": "应标注每种反馈是人工、真值还是学习模块，并保留技能库覆盖范围和重试预算；不能将含人工反馈系统称为全自主。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": null, "codeReleaseScope": "未独立核实" }, { "id": "arxiv-2207.04429", "title": "LM-Nav: Robotic Navigation with Large Pre-Trained Models of Language, Vision, and Action", "date": "2022-07-10", "url": "https://arxiv.org/abs/2207.04429v2", "titleZh": "LM-Nav：组合预训练语言、视觉与动作模型进行机器人导航", "abstractZh": "GPT-3从指令提取地标，CLIP将地标关联到先前探索的图像节点，视觉导航模型估计节点连通并执行图搜索所得路线；三个模型权重均冻结。\n真实Jackal在户外遵循自然语言路线，先人工驾驶构图；LLM/VLM查询在远程工作站预计算，低层视觉导航机载运行。", "summary": "GPT-3从指令提取地标，CLIP将地标关联到先前探索的图像节点，视觉导航模型估计节点连通并执行图搜索所得路线；三个模型权重均冻结。", "experimentType": "real", "robots": [ "Clearpath Jackal" ], "tags": [ "导航与建图", "视觉语言动作" ], "limitations": "主要理解地标，忽略慢速行驶等动词和细节要求；低层模型针对Jackal户外导航，依赖预先探索地图。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的预训练模型组合导航研究，适合理解地标图规划的适用范围。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2207.04429v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2207.04429v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/blazejosinski/lm_nav", "note": "已发布高层规划与实验 notebook；非完整实机导航栈；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/blazejosinski/lm_nav", "category": "导航与建图 / 视觉语言动作", "year": 2022, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "没有完整机器人策略训练；提供图规划与语言消融 notebook。 提供样例图和实验获取入口，不是完整机器人导航数据全集。 依赖预训练外部模型和缓存响应，未发布整套低层导航权重。 这份发布主要是图规划和语言模块，不是完整 Jackal 底层控制与视觉导航模型训练栈。 本次未运行训练、复现实验或实机控制。", "directions": [ "导航与建图", "视觉语言动作" ], "robotFilters": [ "Clearpath Jackal" ], "original": { "id": "arxiv-2207.04429", "originalSourceUrl": "https://arxiv.org/pdf/2207.04429v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2207.04429v2", "finalDownloadUrl": "https://arxiv.org/pdf/2207.04429v2", "sha256": "ca454c56aee5ef49666684b7a8daed62233fa12c439656bcd7f6c51265e89ffb", "bytes": 4715102, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:50.474650+00:00", "publicationDate": "2022/07/10", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2207.04429v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2207.04429v2", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "LM-Nav: Robotic Navigation with Large Pre-Trained Models of Language, Vision, and Action", "sourceVersion": "2207.04429v2", "sourceVersionDate": "2022/07/26", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2207.04429v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2207.04429" }, "originalAnalysis": { "id": "arxiv-2207.04429", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2207.04429v2", "sourceVersion": "2207.04429v2", "sectionsRead": [ "PDF第6页选段（提取文本L440–457）", "PDF第7页选段（提取文本L513–530）", "PDF第8页选段（提取文本L610–623）", "PDF第3页选段（提取文本L185–220）", "PDF第9–10页选段（提取文本L707–741）", "PDF第4页选段（提取文本L255–280）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "GPT-3从指令提取地标，CLIP将地标关联到先前探索的图像节点，视觉导航模型估计节点连通并执行图搜索所得路线；三个模型权重均冻结。", "experimentsZh": "真实Jackal在户外遵循自然语言路线，先人工驾驶构图；LLM/VLM查询在远程工作站预计算，低层视觉导航机载运行。", "resultsZh": "展示达到约800米目标的导航和地标歧义消解；规划成功和实际路线效率分别定义，不应把离线路径选择算成端到端自主探索。", "limitationsZh": "主要理解地标，忽略慢速行驶等动词和细节要求；低层模型针对Jackal户外导航，依赖预先探索地图。", "reproductionZh": "保留构图数据、语言提示、CLIP型号和缓存查询，并区分人工建图、远程规划与机载执行各部分。", "experimentType": "real", "robots": [ "Clearpath Jackal" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "GPT-3从指令提取地标，CLIP将地标关联到先前探索的图像节点，视觉导航模型估计节点连通并执行图搜索所得路线；三个模型权重均冻结。" }, { "section": "原文实验与结果相关选段", "note": "真实Jackal在户外遵循自然语言路线，先人工驾驶构图；LLM/VLM查询在远程工作站预计算，低层视觉导航机载运行。 展示达到约800米目标的导航和地标歧义消解；规划成功和实际路线效率分别定义，不应把离线路径选择算成端到端自主探索。" }, { "section": "局限与复现条件", "note": "主要理解地标，忽略慢速行驶等动词和细节要求；低层模型针对Jackal户外导航，依赖预先探索地图。" } ], "originalSha256": "ca454c56aee5ef49666684b7a8daed62233fa12c439656bcd7f6c51265e89ffb", "readingScopeSha256": "bc49588b52ab951e212c311dce44ae4fd98efe39f6c3220ffe1a6f382f89947e", "textExtractionSha256": "70c23745445f518d9b1b76a23dad98b48a3af497da343386fef9475b09be4115" }, "experimentNote": "真实Jackal在户外遵循自然语言路线，先人工驾驶构图；LLM/VLM查询在远程工作站预计算，低层视觉导航机载运行。", "contribution": "GPT-3从指令提取地标，CLIP将地标关联到先前探索的图像节点，视觉导航模型估计节点连通并执行图搜索所得路线；三个模型权重均冻结。", "whyUseful": "保留构图数据、语言提示、CLIP型号和缓存查询，并区分人工建图、远程规划与机载执行各部分。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "仓库 MIT；外部模型/API 和导航数据各自核对。", "codeReleaseScope": "已发布高层规划与实验 notebook；非完整实机导航栈" }, { "id": "arxiv-2206.14176", "title": "DayDreamer: World Models for Physical Robot Learning", "titleZh": "DayDreamer：用于真实机器人学习的世界模型", "shortTitle": "DayDreamer", "date": "2022-06-28", "year": 2022, "url": "https://arxiv.org/abs/2206.14176", "paperUrl": "https://arxiv.org/abs/2206.14176", "projectUrl": "https://danijar.com/project/daydreamer/", "codeUrl": "https://github.com/danijar/daydreamer", "summary": "把DreamerV2的循环随机世界模型用于真实机器人，融合图像/深度/本体输入，从回放中学习潜在动力学，再在想象轨迹上训练actor-critic。异步分离采集与学习，减少真实控制等待；四平台共享算法超参数，但…", "abstractZh": "把DreamerV2的循环随机世界模型用于真实机器人，融合图像/深度/本体输入，从回放中学习潜在动力学，再在想象轨迹上训练actor-critic。异步分离采集与学习，减少真实控制等待；四平台共享算法超参数，但观察、动作与奖励按任务定义。\n直接在Unitree A1、UR5、7DoF xArm和Sphero Ollie上从头训练，分别对照SAC、Rainbow/PPO、DrQv2。A1用本体、20Hz关节动作，机械臂采用离散末端动作和任务结构约束，Sphero以顶视图与稠密距离奖励学习。", "category": "世界模型 / 实机强化学习", "tags": [ "DayDreamer", "世界模型", "强化学习", "运动控制", "操作与抓取" ], "directions": [ "世界模型", "运动控制", "操作与抓取", "强化学习" ], "tier": "classic", "experimentType": "real", "experimentNote": "直接在Unitree A1、UR5、7DoF xArm和Sphero Ollie上从头训练，分别对照SAC、Rainbow/PPO、DrQv2。A1用本体、20Hz关节动作，机械臂采用离散末端动作和任务结构约束，Sphero以顶视图与稠密距离奖励学习。", "robots": [ "Unitree A1", "Universal Robots UR5", "xArm（7自由度）", "Sphero Ollie" ], "robotFilters": [ "Unitree A1", "Universal Robots UR5", "xArm", "Sphero Ollie" ], "robotNote": "原文明确 Unitree A1、UR5、7-DoF XArm 和 Sphero Ollie；未补写未经本轮来源确认的 XArm 厂商或产品后缀。", "codeStatus": "unknown", "status": "代码公开，许可未核实", "trainingStatus": "代码公开，许可未核实", "trainingNote": "官方 TF2 实现含 actor/learner 和机器人配置；本轮未核实仓库统一许可及所有第三方组件授权。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "未核实", "contribution": "把DreamerV2的循环随机世界模型用于真实机器人，融合图像/深度/本体输入，从回放中学习潜在动力学，再在想象轨迹上训练actor-critic。异步分离采集与学习，减少真实控制等待；四平台共享算法超参数，但观察、动作与奖励按任务定义。", "whyUseful": "复现优先核对异步调度、控制频率、保护滤波、奖励检测和各任务重置条件；“同超参数”不能省略环境工程。正文给出开放软件入口，本轮未实机运行，结果应保留各平台不同时间/任务口径。", "limitations": "无需仿真/示范不代表无人介入：A1到训练区边缘要人工移动，机械臂有Z动作限制/自动放物，xArm对象系绳。A1曲线为单次训练内时间分箱波动，非多种子置信区间；长期试错还可能造成硬件磨损。", "caveats": "硬件在线探索有磨损与安全成本；操作任务有环境约束，不能视为通用自主机器人。", "evidence": [ { "url": "https://arxiv.org/abs/2206.14176", "note": "首发日期。" }, { "url": "https://arxiv.org/html/2206.14176v1", "note": "§3：A1、UR5、7-DoF XArm、Sphero Ollie；§5 硬件磨损限制。" }, { "url": "https://github.com/danijar/daydreamer", "note": "官方实机训练基础设施；统一许可未核实。" }, { "url": "https://arxiv.org/pdf/2206.14176", "note": "Fig.4：A1图为单次训练，阴影是时间分箱内标准差。" }, { "url": "https://arxiv.org/pdf/2206.14176", "note": "§3.1–3.3：边界人工介入、Z动作限制、自动放置与系绳。" }, { "url": "https://arxiv.org/pdf/2206.14176", "note": "§5：长期硬件学习磨损和人工维修限制。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2022：把潜在想象式学习从游戏和仿真推进到多种真实机器人。", "freshness": "经典工作", "original": { "id": "arxiv-2206.14176", "originalSourceUrl": "https://arxiv.org/pdf/2206.14176v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2206.14176v1", "finalDownloadUrl": "https://arxiv.org/pdf/2206.14176", "sha256": "2ee10ae0961bef97ca522bb7ee5b7514e7a918a637dedad569b572b18bccb35e", "bytes": 3126052, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:12.926681+00:00", "publicationDate": "2022/06/28", "publicationDateSourceUrl": "https://arxiv.org/abs/2206.14176", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2206.14176", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2206.14176", "pages": 15, "sourceIdentityStatus": "title_match", "sourceTitle": "DayDreamer: World Models for Physical Robot Learning", "sourceVersion": "2206.14176v1", "sourceVersionDate": "2022/06/28", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.234333+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2206.14176" }, "originalAnalysis": { "id": "arxiv-2206.14176", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090143+00:00", "sourceUrl": "https://arxiv.org/pdf/2206.14176", "sectionsRead": [ "§2", "§3.1–3.4", "Figs.4–7", "§5" ], "methodsZh": "把DreamerV2的循环随机世界模型用于真实机器人，融合图像/深度/本体输入，从回放中学习潜在动力学，再在想象轨迹上训练actor-critic。异步分离采集与学习，减少真实控制等待；四平台共享算法超参数，但观察、动作与奖励按任务定义。", "experimentsZh": "直接在Unitree A1、UR5、7DoF xArm和Sphero Ollie上从头训练，分别对照SAC、Rainbow/PPO、DrQv2。A1用本体、20Hz关节动作，机械臂采用离散末端动作和任务结构约束，Sphero以顶视图与稠密距离奖励学习。", "resultsZh": "作者报告A1约一小时学翻身站立行走，随后十分钟适应推扰；UR5八小时约2.5物体/分钟，xArm十小时3.1物体/分钟，Sphero两小时可导航。Sphero与DrQv2相近，并非全部平台均显著胜出。", "limitationsZh": "无需仿真/示范不代表无人介入：A1到训练区边缘要人工移动，机械臂有Z动作限制/自动放物，xArm对象系绳。A1曲线为单次训练内时间分箱波动，非多种子置信区间；长期试错还可能造成硬件磨损。", "reproductionZh": "复现优先核对异步调度、控制频率、保护滤波、奖励检测和各任务重置条件；“同超参数”不能省略环境工程。正文给出开放软件入口，本轮未实机运行，结果应保留各平台不同时间/任务口径。", "experimentType": "real", "robots": [ "Unitree A1", "Universal Robots UR5", "xArm（7自由度）", "Sphero Ollie" ], "evidenceNotes": [ { "section/page": "Fig.4", "note": "A1图为单次训练，阴影是时间分箱内标准差。" }, { "section/page": "§3.1–3.3", "note": "边界人工介入、Z动作限制、自动放置与系绳。" }, { "section/page": "§5", "note": "长期硬件学习磨损和人工维修限制。" } ] }, "analysisVerifiedAt": "2026-10-04T13:50:35.090143+00:00" }, { "id": "arxiv-2205.09991", "title": "Planning with Diffusion for Flexible Behavior Synthesis", "date": "2022-05-20", "url": "https://arxiv.org/abs/2205.09991v2", "titleZh": "Diffuser：用扩散规划合成灵活行为", "abstractZh": "把整段状态-动作轨迹建模为扩散去噪过程，以目标条件或引导函数影响采样；通过局部时序卷积多轮去噪生成全局连贯计划。\n在Maze2D/Multi2D、仿真堆块和D4RL连续控制任务比较长时序规划、测试时目标变更及不同质量离线数据。", "summary": "把整段状态-动作轨迹建模为扩散去噪过程，以目标条件或引导函数影响采样；通过局部时序卷积多轮去噪生成全局连贯计划。", "experimentType": "sim", "robots": [], "tags": [ "世界模型", "强化学习", "基础理论" ], "limitations": "迭代采样慢是原文明确瓶颈；验证以仿真状态轨迹为主，不能推导图像感知或真实接触控制同样有效。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "foundation", "tierNote": "编辑分类：扩散轨迹规划的重要方法入口，提供理解后续动作扩散、目标引导与采样开销的基础概念。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2205.09991v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2205.09991v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/jannerm/diffuser", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/jannerm/diffuser", "category": "世界模型 / 强化学习", "year": 2022, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "扩散模型和价值函数训练代码已发布。 通过 D4RL 获取离线数据；数据集有独立条件。 作者提供预训练模型、价值函数及规划结果。 主分支是仿真运动任务；KUKA 分支并不构成实机验证。不要与 Diffusion Policy 或 Hugging Face diffusers 库混淆。 本次未运行训练、复现实验或实机控制。", "directions": [ "世界模型", "强化学习", "基础理论" ], "robotFilters": [], "original": { "id": "arxiv-2205.09991", "originalSourceUrl": "https://arxiv.org/pdf/2205.09991v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2205.09991v2", "finalDownloadUrl": "https://arxiv.org/pdf/2205.09991v2", "sha256": "aafe149adb5476e8b8b061c02d26824c42fb14a71ce8ae37b1a0e375452d7a1f", "bytes": 10498721, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:01:10.785629+00:00", "publicationDate": "2022/05/20", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2205.09991v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2205.09991v2", "pages": 14, "sourceIdentityStatus": "title_match", "sourceTitle": "Planning with Diffusion for Flexible Behavior Synthesis", "sourceVersion": "2205.09991v2", "sourceVersionDate": "2022/12/21", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2205.09991v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2205.09991" }, "originalAnalysis": { "id": "arxiv-2205.09991", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2205.09991v2", "sourceVersion": "2205.09991v2", "sectionsRead": [ "PDF第3页选段（提取文本L273–290）", "PDF第6页选段（提取文本L642–655）", "PDF第6页选段（提取文本L660–677）", "PDF第9页选段（提取文本L1144–1171）", "PDF第6–7页选段（提取文本L702–745）", "PDF第8页选段（提取文本L1049–1087）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "把整段状态-动作轨迹建模为扩散去噪过程，以目标条件或引导函数影响采样；通过局部时序卷积多轮去噪生成全局连贯计划。", "experimentsZh": "在Maze2D/Multi2D、仿真堆块和D4RL连续控制任务比较长时序规划、测试时目标变更及不同质量离线数据。", "resultsZh": "可在不重训的情况下改变规划目标；用上一时刻计划热启动并减少去噪步数，可在文中任务中以有限性能损失降低规划开销。", "limitationsZh": "迭代采样慢是原文明确瓶颈；验证以仿真状态轨迹为主，不能推导图像感知或真实接触控制同样有效。", "reproductionZh": "固定离线数据版本、引导强度、规划长度、去噪步数和重规划节奏，分别报告归一化回报及实际规划耗时。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "把整段状态-动作轨迹建模为扩散去噪过程，以目标条件或引导函数影响采样；通过局部时序卷积多轮去噪生成全局连贯计划。" }, { "section": "原文实验与结果相关选段", "note": "在Maze2D/Multi2D、仿真堆块和D4RL连续控制任务比较长时序规划、测试时目标变更及不同质量离线数据。 可在不重训的情况下改变规划目标；用上一时刻计划热启动并减少去噪步数，可在文中任务中以有限性能损失降低规划开销。" }, { "section": "局限与复现条件", "note": "迭代采样慢是原文明确瓶颈；验证以仿真状态轨迹为主，不能推导图像感知或真实接触控制同样有效。" } ], "originalSha256": "aafe149adb5476e8b8b061c02d26824c42fb14a71ce8ae37b1a0e375452d7a1f", "readingScopeSha256": "96f205f255aeb18ff1691b931f90ddae761b0126f4338d80cc144c2e04cef1b9", "textExtractionSha256": "a22cdaad52b00e9bde3cbcddbfdca4f5e7e2b50a396cc9208563b4267414fed9" }, "experimentNote": "在Maze2D/Multi2D、仿真堆块和D4RL连续控制任务比较长时序规划、测试时目标变更及不同质量离线数据。", "contribution": "把整段状态-动作轨迹建模为扩散去噪过程，以目标条件或引导函数影响采样；通过局部时序卷积多轮去噪生成全局连贯计划。", "whyUseful": "固定离线数据版本、引导强度、规划长度、去噪步数和重规划节奏，分别报告归一化回报及实际规划耗时。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "主代码 MIT；D4RL、仿真依赖、数据和下载检查点需要分别核对。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2205.01906", "title": "ASE: Large-Scale Reusable Adversarial Skill Embeddings for Physically Simulated Characters", "titleZh": "ASE：物理仿真角色的大规模可复用对抗技能嵌入", "shortTitle": "ASE", "date": "2022-05-04", "datePrecision": "day", "year": 2022, "url": "https://arxiv.org/abs/2205.01906", "paperUrl": "https://arxiv.org/abs/2205.01906", "codeUrl": "https://github.com/nv-tlabs/ASE", "summary": "ASE以对抗模仿约束动作风格，并最大化技能潜变量与状态转移互信息，使无标签动作库形成可复用的连续技能空间。低层策略以单位球面技能控制关节，高层学习任务到技能选择，沿用冻结判别器作运动先验以减少不自然切换。", "abstractZh": "ASE以对抗模仿约束动作风格，并最大化技能潜变量与状态转移互信息，使无标签动作库形成可复用的连续技能空间。低层策略以单位球面技能控制关节，高层学习任务到技能选择，沿用冻结判别器作运动先验以减少不自然切换。\n在Isaac Gym中用37自由度持剑盾人形角色，187段约30分钟动作训练低层；4096并行环境、单V100约十天及百亿样本。五种下游任务评测三套预训练低层，每套4096回合，另检查技能覆盖、切换和跌倒恢复。", "category": "运动控制 / 技能表示", "tags": [ "技能嵌入", "对抗模仿", "分层强化学习", "运动预训练" ], "directions": [ "运动控制", "基础理论", "强化学习", "模仿学习" ], "tier": "foundation", "experimentType": "sim", "robots": [], "robotNote": "物理仿真人形角色，未核实任何实体机器人实验。", "codeStatus": "open", "status": "作者预训练与评估代码公开；旧代码库已标注弃用。", "trainingNote": "Isaac Gym中预训练低层技能，再服务下游任务；作者现推荐MimicKit，项目动作数据限非商业用途。", "contribution": "ASE以对抗模仿约束动作风格，并最大化技能潜变量与状态转移互信息，使无标签动作库形成可复用的连续技能空间。低层策略以单位球面技能控制关节，高层学习任务到技能选择，沿用冻结判别器作运动先验以减少不自然切换。", "whyUseful": "需获得对应动作资产和重定向骨架，保留30Hz低层、6Hz高层及120Hz物理频率，复现潜变量归一化、互信息与判别器。报告预训练成本和下游增量样本，未运行实验。", "limitations": "只有物理动画仿真，无人形机器人硬件或传感器噪声验证。预训练费用高，技能受动作库和身体形态约束，风格/任务奖励仍手工加权；自然性主要依视觉和先验，不能当成硬件安全证据。", "evidence": [ { "url": "https://arxiv.org/abs/2205.01906", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://research.nvidia.com/labs/toronto-ai/ASE/", "note": "团队项目页说明方法、下游使用与动作数据非商业限制。" }, { "url": "https://github.com/nv-tlabs/ASE", "note": "作者README含预训练、测试命令及弃用提醒。" }, { "url": "https://arxiv.org/pdf/2205.01906", "note": "§8–10.1：37DoF虚拟角色、187片段、百亿样本和单V100十天。" }, { "url": "https://arxiv.org/pdf/2205.01906", "note": "Table 1; §10.4：三模型×4096回合、归一化回报及从零策略不自然的取舍。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2022：从单任务运动模仿迈向大规模预训练且可复用的技能空间。", "projectUrl": null, "robotFilters": [], "original": { "id": "arxiv-2205.01906", "originalSourceUrl": "https://arxiv.org/pdf/2205.01906v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2205.01906v2", "finalDownloadUrl": "https://arxiv.org/pdf/2205.01906", "sha256": "e4151dfddd0f79ef496092b58c10ef67b519e9a5be3b4d5b2067c575124050c7", "bytes": 24478689, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:18.131012+00:00", "publicationDate": "2022/05/04", "publicationDateSourceUrl": "https://arxiv.org/abs/2205.01906", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2205.01906", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2205.01906", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "ASE: Large-Scale Reusable Adversarial Skill Embeddings for Physically Simulated Characters", "sourceVersion": "2205.01906v2", "sourceVersionDate": "2022/05/05", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.251311+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2205.01906" }, "originalAnalysis": { "id": "arxiv-2205.01906", "sourceUrl": "https://arxiv.org/pdf/2205.01906", "sectionsRead": [ "§5–7", "§8–10", "Table 1", "§10.4–10.6" ], "methodsZh": "ASE以对抗模仿约束动作风格，并最大化技能潜变量与状态转移互信息，使无标签动作库形成可复用的连续技能空间。低层策略以单位球面技能控制关节，高层学习任务到技能选择，沿用冻结判别器作运动先验以减少不自然切换。", "experimentsZh": "在Isaac Gym中用37自由度持剑盾人形角色，187段约30分钟动作训练低层；4096并行环境、单V100约十天及百亿样本。五种下游任务评测三套预训练低层，每套4096回合，另检查技能覆盖、切换和跌倒恢复。", "resultsZh": "作者报告到达、速度、转向、位置和打击归一化回报分别约0.75、0.93、0.90、0.45、0.82。从零训练常有更高任务回报却产生不自然动作；ASE收益是自然风格与技能复用，并非所有回报最高。覆盖和技能发现消融显示减轻模式塌缩。", "limitationsZh": "只有物理动画仿真，无人形机器人硬件或传感器噪声验证。预训练费用高，技能受动作库和身体形态约束，风格/任务奖励仍手工加权；自然性主要依视觉和先验，不能当成硬件安全证据。", "reproductionZh": "需获得对应动作资产和重定向骨架，保留30Hz低层、6Hz高层及120Hz物理频率，复现潜变量归一化、互信息与判别器。报告预训练成本和下游增量样本，未运行实验。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "§8–10.1", "note": "37DoF虚拟角色、187片段、百亿样本和单V100十天。" }, { "section/page": "Table 1; §10.4", "note": "三模型×4096回合、归一化回报及从零策略不自然的取舍。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "在Isaac Gym中用37自由度持剑盾人形角色，187段约30分钟动作训练低层；4096并行环境、单V100约十天及百亿样本。五种下游任务评测三套预训练低层，每套4096回合，另检查技能覆盖、切换和跌倒恢复。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2204.01691", "title": "Do As I Can, Not As I Say: Grounding Language in Robotic Affordances", "titleZh": "按我能做的来：利用机器人可供性约束语言规划", "shortTitle": "SayCan", "date": "2022-04-04", "year": 2022, "url": "https://arxiv.org/abs/2204.01691", "codeUrl": "https://github.com/google-research/google-research/tree/master/saycan", "category": "具身推理与规划", "tags": [ "LLM", "可供性", "技能组合", "长时序" ], "tier": "foundation", "summary": "SayCan对有限技能库的自然语言描述计算LLM条件概率，同时用技能价值函数估计当前场景能否成功；两者相乘后选择下一技能，执行后重新打分直至终止。强化学习的价值函数提供可执行性，LLM提供对长指令的语义步骤偏好…", "abstractZh": "SayCan对有限技能库的自然语言描述计算LLM条件概率，同时用技能价值函数估计当前场景能否成功；两者相乘后选择下一技能，执行后重新打分直至终止。强化学习的价值函数提供可执行性，LLM提供对长指令的语义步骤偏好。\nEveryday Robots七自由度移动机械臂，在仿办公厨房及真实办公厨房使用15物体、5个语义位置，101条指令涵盖七族。主要LLM为540B PaLM，分别由人工判规划是否正确与完整真实执行是否完成，并做去可供性、换LLM等消融。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "real", "robots": [ "Everyday Robots移动机械臂" ], "codeStatus": "open", "status": "官方开放桌面仿真版本；不是完整厨房实机系统。", "trainingNote": "Google Research 仓库 Apache-2.0；实机低层技能包含示范学习与仿真强化学习，完整 PaLM 配置不等同于开源演示。", "whyUseful": "先复现技能描述的概率归一化、终止候选和价值校准，再在固定技能库比较语言-only、可供性-only及乘积选择。保留实际执行后的重规划记录；读者应把技能学习成本与上层零样本指令规划区分。", "contribution": "SayCan对有限技能库的自然语言描述计算LLM条件概率，同时用技能价值函数估计当前场景能否成功；两者相乘后选择下一技能，执行后重新打分直至终止。强化学习的价值函数提供可执行性，LLM提供对长指令的语义步骤偏好。", "limitations": "这不是LLM直接生成连续控制，也不能凭新语言创造未训练技能；系统受技能范围、价值函数错误及语言偏差共同限制。两厨房及固定对象位置范围有限，人工计划判定与动作成功需要分开阅读。", "caveats": "实验标签按论文主要厨房实机评测；公开仿真演示另行说明。", "evidence": [ { "url": "https://arxiv.org/abs/2204.01691", "note": "初版日期，v2 加入 PaLM 与仿真开源版本。" }, { "url": "https://arxiv.org/html/2204.01691v2", "note": "第 5 节明确 Everyday Robots 硬件与真实厨房评测。" }, { "url": "https://say-can.github.io/", "note": "官方开源范围限定为模拟桌面环境。" }, { "url": "https://github.com/google-research/google-research/tree/master/saycan", "note": "实际 notebook 实现。" }, { "url": "https://github.com/google-research/google-research/blob/master/LICENSE", "note": "仓库 Apache-2.0 许可证。" }, { "url": "https://arxiv.org/pdf/2204.01691", "note": "3–4：语言概率乘技能价值与逐步重规划。" }, { "url": "https://arxiv.org/pdf/2204.01691", "note": "5 Tables 1–3; 8：101指令、两厨房、84/74与技能瓶颈。" }, { "url": "https://arxiv.org/pdf/2204.01691v2", "note": "3–4：语言概率乘技能价值与逐步重规划。" }, { "url": "https://arxiv.org/pdf/2204.01691v2", "note": "5 Tables 1–3; 8：101指令、两厨房、84/74与技能瓶颈。" } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "其他机器人研究" ], "robotFilters": [], "verificationNote": "已核硬件、评测与开源范围；未将仿真 notebook 视为完整实机复现。", "original": { "id": "arxiv-2204.01691", "originalSourceUrl": "https://arxiv.org/pdf/2204.01691v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2204.01691v2", "finalDownloadUrl": "https://arxiv.org/pdf/2204.01691", "sha256": "48c22d36a044bbbd34e640beb1db4fb4ba57e74d97fefa0fb121d55c31282312", "bytes": 12827088, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:17.962799+00:00", "publicationDate": "2022/04/04", "publicationDateSourceUrl": "https://arxiv.org/abs/2204.01691", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2204.01691", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2204.01691", "pages": 34, "sourceIdentityStatus": "title_match", "sourceTitle": "Do As I Can, Not As I Say: Grounding Language in Robotic Affordances", "sourceVersion": "2204.01691v2", "sourceVersionDate": "2022/08/16", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.262221+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2204.01691" }, "originalAnalysis": { "id": "arxiv-2204.01691", "sourceUrl": "https://arxiv.org/pdf/2204.01691v2", "sectionsRead": [ "3–4", "5.1–5.3", "8", "Tables 1–3" ], "methodsZh": "SayCan对有限技能库的自然语言描述计算LLM条件概率，同时用技能价值函数估计当前场景能否成功；两者相乘后选择下一技能，执行后重新打分直至终止。强化学习的价值函数提供可执行性，LLM提供对长指令的语义步骤偏好。", "experimentsZh": "Everyday Robots七自由度移动机械臂，在仿办公厨房及真实办公厨房使用15物体、5个语义位置，101条指令涵盖七族。主要LLM为540B PaLM，分别由人工判规划是否正确与完整真实执行是否完成，并做去可供性、换LLM等消融。", "resultsZh": "仿办公厨房规划84%、执行74%；真实厨房分别下降3和14个百分点，显示低层技能可靠性对端到端结果有明显影响。长指令可通过技能组合完成，LLM升级改善规划，但动作执行仍必须落在既有技能集合内。", "limitationsZh": "这不是LLM直接生成连续控制，也不能凭新语言创造未训练技能；系统受技能范围、价值函数错误及语言偏差共同限制。两厨房及固定对象位置范围有限，人工计划判定与动作成功需要分开阅读。", "reproductionZh": "先复现技能描述的概率归一化、终止候选和价值校准，再在固定技能库比较语言-only、可供性-only及乘积选择。保留实际执行后的重规划记录；读者应把技能学习成本与上层零样本指令规划区分。", "experimentType": "real", "robots": [ "Everyday Robots移动机械臂" ], "corrections": [ "零样本指的是高层指令组合，底层技能和价值函数经过大量训练。" ], "evidenceNotes": [ { "note": "语言概率乘技能价值与逐步重规划。", "section": "3–4" }, { "note": "101指令、两厨房、84/74与技能瓶颈。", "section": "5 Tables 1–3; 8" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2204.01691v2", "originalSha256": "48c22d36a044bbbd34e640beb1db4fb4ba57e74d97fefa0fb121d55c31282312" }, "experimentNote": "Everyday Robots七自由度移动机械臂，在仿办公厨房及真实办公厨房使用15物体、5个语义位置，101条指令涵盖七族。主要LLM为540B PaLM，分别由人工判规划是否正确与完整真实执行是否完成，并做去可供性、换LLM等消融。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2203.04955", "title": "Temporal Difference Learning for Model Predictive Control", "titleZh": "用于模型预测控制的时序差分学习", "shortTitle": "TD-MPC", "date": "2022-03-09", "year": 2022, "url": "https://arxiv.org/abs/2203.04955", "paperUrl": "https://arxiv.org/abs/2203.04955", "projectUrl": null, "codeUrl": "https://github.com/nicklashansen/tdmpc", "summary": "TD-MPC联合学习编码器、确定性潜动力学、奖励、Q值和辅助策略，用多步奖励误差、TD目标及潜状态一致性训练，不重建像素。执行时在潜空间采样短动作序列，将预测短期奖励和末端Q值相加，再按高分轨迹更新分布；辅助策…", "abstractZh": "TD-MPC联合学习编码器、确定性潜动力学、奖励、Q值和辅助策略，用多步奖励误差、TD目标及潜状态一致性训练，不重建像素。执行时在潜空间采样短动作序列，将预测短期奖励和末端Q值相加，再按高分轨迹更新分布；辅助策略提供部分候选轨迹。\n共评估DMControl及Meta-World v2的92项仿真任务，覆盖状态、像素、稀疏奖励、多任务和本体加视觉输入。状态曲线通常五种子，像素100k表为十次；比较SAC、LOOP、像素RL与带真模拟器的MPC，并移除一致性或更换重建/对比目标。", "category": "世界模型 / 模型预测控制", "tags": [ "TD-MPC", "世界模型", "强化学习" ], "directions": [ "世界模型", "强化学习" ], "tier": "classic", "experimentType": "sim", "experimentNote": "共评估DMControl及Meta-World v2的92项仿真任务，覆盖状态、像素、稀疏奖励、多任务和本体加视觉输入。状态曲线通常五种子，像素100k表为十次；比较SAC、LOOP、像素RL与带真模拟器的MPC，并移除一致性或更换重建/对比目标。", "robots": [], "robotFilters": [], "robotNote": "无实机；仿真形态不等同于已验证的商业机器人型号。", "codeStatus": "open", "status": "训练代码已开源", "trainingStatus": "训练代码已开源", "trainingNote": "官方 PyTorch 训练实现与配置公开，MIT 许可；旧仿真依赖需单独处理。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "MIT（代码）", "contribution": "TD-MPC联合学习编码器、确定性潜动力学、奖励、Q值和辅助策略，用多步奖励误差、TD目标及潜状态一致性训练，不重建像素。执行时在潜空间采样短动作序列，将预测短期奖励和末端Q值相加，再按高分轨迹更新分布；辅助策略提供部分候选轨迹。", "whyUseful": "标准规划视野5步、512轨迹加5%策略候选、64精英，通常六轮优化；同时固定优先回放、目标网络、动作重复和环境版本。复现需区分环境步与决策步，分别报告样本量、达到阈值时间及单次推理预算。", "limitations": "任务相关潜表示不保证对无关新任务通用，确定性模型未显式校准不确定性；在线优化仍有延迟。论文全部为仿真，不应把后续TD-MPC2或真实机器人工作归入本篇；部分基线数字沿用原作者发布结果。", "caveats": "需要在线奖励反馈与规划计算；原论文为仿真基准，未验证真实机器人。", "evidence": [ { "url": "https://arxiv.org/abs/2203.04955", "note": "首发日期与任务范围。" }, { "url": "https://arxiv.org/html/2203.04955v2", "note": "任务导向潜动力学、TD 学习与仿真实验。" }, { "url": "https://github.com/nicklashansen/tdmpc", "note": "官方原始 PyTorch 代码与 MIT。" }, { "url": "https://arxiv.org/pdf/2203.04955v2", "note": "3/4 式3、7–11：Short rollouts plus terminal Q; reward/value/latent-consistency training; deterministic latent components." }, { "url": "https://arxiv.org/pdf/2203.04955v2", "note": "5 表1：Image-based100k benchmark uses10 runs; task-specific gains, not uniform superiority." }, { "url": "https://arxiv.org/pdf/2203.04955v2", "note": "5 表2：Walker solve threshold940;0.47h TD-MPC,7.72h LOOP,0.41h SAC; per500k TD-MPC5.60h vs SAC1.41h." }, { "url": "https://arxiv.org/pdf/2203.04955v2", "note": "附录F 表4/7：H5;512 samples+5% policy;64 elites;action repeat task-dependent." } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2022：把短时域模型规划与长期 TD 价值结合为实用连续控制基线。", "freshness": "经典工作", "original": { "id": "arxiv-2203.04955", "originalSourceUrl": "https://arxiv.org/pdf/2203.04955v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2203.04955v2", "finalDownloadUrl": "https://arxiv.org/pdf/2203.04955", "sha256": "b80e9fc9eaebfa43ad30641ebb0327346998e604cd7421ae44701ef373b501aa", "bytes": 12584890, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:23.007731+00:00", "publicationDate": "2022/03/09", "publicationDateSourceUrl": "https://arxiv.org/abs/2203.04955", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2203.04955", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2203.04955", "pages": 20, "sourceIdentityStatus": "title_match", "sourceTitle": "Temporal Difference Learning for Model Predictive Control", "sourceVersion": "2203.04955v2", "sourceVersionDate": "2022/07/19", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.271981+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2203.04955" }, "originalAnalysis": { "id": "arxiv-2203.04955", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739400+00:00", "sourceUrl": "https://arxiv.org/pdf/2203.04955v2", "sectionsRead": [ "3 MPPI式短程规划与策略引导", "4 TOLD奖励、TD值与潜状态一致性", "5 92任务实验、表1/2和图3–6", "附录F 超参与动作重复" ], "methodsZh": "TD-MPC联合学习编码器、确定性潜动力学、奖励、Q值和辅助策略，用多步奖励误差、TD目标及潜状态一致性训练，不重建像素。执行时在潜空间采样短动作序列，将预测短期奖励和末端Q值相加，再按高分轨迹更新分布；辅助策略提供部分候选轨迹。", "experimentsZh": "共评估DMControl及Meta-World v2的92项仿真任务，覆盖状态、像素、稀疏奖励、多任务和本体加视觉输入。状态曲线通常五种子，像素100k表为十次；比较SAC、LOOP、像素RL与带真模拟器的MPC，并移除一致性或更换重建/对比目标。", "resultsZh": "作者在单RTX3090上报告Walker Walk达到940回报需0.47小时，LOOP为7.72小时，SAC为0.41小时；因此优势是样本与规划效率的折中，并非每步计算比SAC更少。像素100k中Finger Spin为943±59，但Cheetah Run为222±88，结果有任务差异。", "limitationsZh": "任务相关潜表示不保证对无关新任务通用，确定性模型未显式校准不确定性；在线优化仍有延迟。论文全部为仿真，不应把后续TD-MPC2或真实机器人工作归入本篇；部分基线数字沿用原作者发布结果。", "reproductionZh": "标准规划视野5步、512轨迹加5%策略候选、64精英，通常六轮优化；同时固定优先回放、目标网络、动作重复和环境版本。复现需区分环境步与决策步，分别报告样本量、达到阈值时间及单次推理预算。", "evidenceNotes": [ { "section/page": "3/4 式3、7–11", "note": "Short rollouts plus terminal Q; reward/value/latent-consistency training; deterministic latent components." }, { "section/page": "5 表1", "note": "Image-based100k benchmark uses10 runs; task-specific gains, not uniform superiority." }, { "section/page": "5 表2", "note": "Walker solve threshold940;0.47h TD-MPC,7.72h LOOP,0.41h SAC; per500k TD-MPC5.60h vs SAC1.41h." }, { "section/page": "附录F 表4/7", "note": "H5;512 samples+5% policy;64 elites;action repeat task-dependent." } ], "experimentType": "sim", "robots": [], "corrections": [ "原始TD-MPC实验是仿真；不要回填后续版本实机结果。" ], "sourceVersion": "2203.04955v2", "originalSha256": "b80e9fc9eaebfa43ad30641ebb0327346998e604cd7421ae44701ef373b501aa" }, "analysisVerifiedAt": "2026-10-04T13:47:58.739400+00:00" }, { "id": "arxiv-2112.12130", "title": "NICE-SLAM: Neural Implicit Scalable Encoding for SLAM", "date": "2021-12-22", "url": "https://arxiv.org/abs/2112.12130v2", "titleZh": "NICE-SLAM：可扩展的分层神经隐式SLAM", "abstractZh": "用粗中细多层特征网格及预训练占据解码器表示几何，另建颜色网格，通过可微渲染联合优化场景与相机位姿。\n在Replica、ScanNet、TUM RGB-D、Co-Fusion和自采公寓序列上比较跟踪、重建、内存与运行时间，输入为RGB-D。", "summary": "用粗中细多层特征网格及预训练占据解码器表示几何，另建颜色网格，通过可微渲染联合优化场景与相机位姿。", "experimentType": "data", "robots": [], "tags": [ "导航与建图" ], "limitations": "粗层表示限制可预测的空间尺度，原系统没有回环闭合；神经系统与成熟传统SLAM仍存在性能差距。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的分层神经隐式SLAM研究，适合与单MLP及高斯地图对照。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2112.12130v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2112.12130v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/cvg/nice-slam", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/cvg/nice-slam", "category": "导航与建图", "year": 2021, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "包含逐场景在线优化和运行代码，不是机器人行为策略训练。 脚本提供 demo/重建结果与部分数据链接，ScanNet 需按官方流程获取。 仓库含 pretrained 目录；不代表全部上游预训练数据和脚本已开放。 仓库 iMAP* 是作者重实现和调整版本，不能替代原 iMAP 官方代码；场景结果与输入深度质量有关。 本次未运行训练、复现实验或实机控制。", "directions": [ "导航与建图" ], "robotFilters": [], "original": { "id": "arxiv-2112.12130", "originalSourceUrl": "https://arxiv.org/pdf/2112.12130v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2112.12130v2", "finalDownloadUrl": "https://arxiv.org/pdf/2112.12130v2", "sha256": "734cc95a355c07c035696ef7610161522a64af4e9f86d99c2ab3eb1154eb05c6", "bytes": 27574309, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:38.532507+00:00", "publicationDate": "2021/12/22", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2112.12130v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2112.12130v2", "pages": 17, "sourceIdentityStatus": "title_match", "sourceTitle": "NICE-SLAM: Neural Implicit Scalable Encoding for SLAM", "sourceVersion": "2112.12130v2", "sourceVersionDate": "2022/04/21", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2112.12130v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2112.12130" }, "originalAnalysis": { "id": "arxiv-2112.12130", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2112.12130v2", "sourceVersion": "2112.12130v2", "sectionsRead": [ "PDF第4页选段（提取文本L803–820）", "PDF第5页选段（提取文本L852–869）", "PDF第6页选段（提取文本L1101–1114）", "PDF第7页选段（提取文本L1139–1156）", "PDF第3页选段（提取文本L671–708）", "PDF第8页选段（提取文本L1287–1299）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "用粗中细多层特征网格及预训练占据解码器表示几何，另建颜色网格，通过可微渲染联合优化场景与相机位姿。", "experimentsZh": "在Replica、ScanNet、TUM RGB-D、Co-Fusion和自采公寓序列上比较跟踪、重建、内存与运行时间，输入为RGB-D。", "resultsZh": "相较iMAP，分层局部表示提升几何质量和大场景扩展能力；论文在相同像素采样下报告跟踪和建图更快。", "limitationsZh": "粗层表示限制可预测的空间尺度，原系统没有回环闭合；神经系统与成熟传统SLAM仍存在性能差距。", "reproductionZh": "固定场景边界、各级网格分辨率、预训练解码器及采样预算；不能把局部连续跟踪视为带全局回环的完整定位能力。", "experimentType": "data", "robots": [], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "用粗中细多层特征网格及预训练占据解码器表示几何，另建颜色网格，通过可微渲染联合优化场景与相机位姿。" }, { "section": "原文实验与结果相关选段", "note": "在Replica、ScanNet、TUM RGB-D、Co-Fusion和自采公寓序列上比较跟踪、重建、内存与运行时间，输入为RGB-D。 相较iMAP，分层局部表示提升几何质量和大场景扩展能力；论文在相同像素采样下报告跟踪和建图更快。" }, { "section": "局限与复现条件", "note": "粗层表示限制可预测的空间尺度，原系统没有回环闭合；神经系统与成熟传统SLAM仍存在性能差距。" } ], "originalSha256": "734cc95a355c07c035696ef7610161522a64af4e9f86d99c2ab3eb1154eb05c6", "readingScopeSha256": "eadbcf9e5a8b3f5a12df3a8d5bd58022d9e40dbd53e873b3e60f078efe76c87f", "textExtractionSha256": "1a2174a3e3aa41b856ba570302a2a6c35476ed04b8fad2be85df643d79bc27e3" }, "experimentNote": "在Replica、ScanNet、TUM RGB-D、Co-Fusion和自采公寓序列上比较跟踪、重建、内存与运行时间，输入为RGB-D。", "contribution": "用粗中细多层特征网格及预训练占据解码器表示几何，另建颜色网格，通过可微渲染联合优化场景与相机位姿。", "whyUseful": "固定场景边界、各级网格分辨率、预训练解码器及采样预算；不能把局部连续跟踪视为带全局回环的完整定位能力。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "仓库 Apache-2.0；ScanNet、Replica、TUM 与其他资产有独立使用条款。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2112.03227", "title": "CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks", "date": "2021-12-06", "url": "https://arxiv.org/abs/2112.03227v4", "titleZh": "CALVIN：长时序语言条件机器人操作基准", "abstractZh": "四套视觉布局中的34项操作技能组合成语言指令链，提供多模态观测和自由交互示教，用来测试语言落地与连续控制。\n区分单环境、多个已见环境和三环境训练/第四环境零样本测试；除单任务外还评估连续五条指令完成情况。", "summary": "四套视觉布局中的34项操作技能组合成语言指令链，提供多模态观测和自由交互示教，用来测试语言落地与连续控制。", "experimentType": "sim", "robots": [ "Franka Emika Panda（仿真）" ], "tags": [ "数据集与基准", "模仿学习", "视觉语言动作" ], "limitations": "物体被简化为基础形状，属于仿真桌面场景；语言、颜色及域偏移会引起失败，不等于真实家务能力。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "foundation", "tierNote": "编辑分类：语言条件长时序操作的常用基准与数据入口，适合先理解任务链和跨环境评估协议。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2112.03227v4", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2112.03227v4", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/mees/calvin", "note": "已核验官方代码与文档；未运行复现；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/mees/calvin", "category": "数据集与基准 / 模仿学习", "year": 2021, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "README已核实D/ABC/ABCD数据集拆分、基线training.py和评估入口；依赖版本与数据下载分开核验。 本次未运行训练、复现实验或实机控制。", "directions": [ "数据集与基准", "模仿学习", "视觉语言动作" ], "robotFilters": [ "Franka Emika Panda（仿真）" ], "original": { "id": "arxiv-2112.03227", "originalSourceUrl": "https://arxiv.org/pdf/2112.03227v4", "versionedOriginalUrl": "https://arxiv.org/pdf/2112.03227v4", "finalDownloadUrl": "https://arxiv.org/pdf/2112.03227v4", "sha256": "c8ad4741782fa8fb773377fbf61b8772b86574f255599a229d3f6ad4d878c60e", "bytes": 2092247, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T14:59:56.955250+00:00", "publicationDate": "2021/12/06", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2112.03227v4", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2112.03227v4", "pages": 10, "sourceIdentityStatus": "title_match", "sourceTitle": "CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks", "sourceVersion": "2112.03227v4", "sourceVersionDate": "2022/07/13", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2112.03227v4", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2112.03227" }, "originalAnalysis": { "id": "arxiv-2112.03227", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2112.03227v4", "sourceVersion": "2112.03227v4", "sectionsRead": [ "PDF第3页选段（提取文本L221–234）", "PDF第4–5页选段（提取文本L405–422）", "PDF第6–7页选段（提取文本L737–750）", "PDF第5页选段（提取文本L427–467）", "PDF第7页选段（提取文本L914–948）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "四套视觉布局中的34项操作技能组合成语言指令链，提供多模态观测和自由交互示教，用来测试语言落地与连续控制。", "experimentsZh": "区分单环境、多个已见环境和三环境训练/第四环境零样本测试；除单任务外还评估连续五条指令完成情况。", "resultsZh": "论文中的MCIL基线在同环境五指令链上的成功率仅0.08%，展示长时序与跨环境泛化的难度；不是当前方法的上限。", "limitationsZh": "物体被简化为基础形状，属于仿真桌面场景；语言、颜色及域偏移会引起失败，不等于真实家务能力。", "reproductionZh": "严格记录ABC→D等训练/测试组合、传感器配置、动作表示及指令链采样，避免把单技能成功率与整链成功率混合。", "experimentType": "sim", "robots": [ "Franka Emika Panda（仿真）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "四套视觉布局中的34项操作技能组合成语言指令链，提供多模态观测和自由交互示教，用来测试语言落地与连续控制。" }, { "section": "原文实验与结果相关选段", "note": "区分单环境、多个已见环境和三环境训练/第四环境零样本测试；除单任务外还评估连续五条指令完成情况。 论文中的MCIL基线在同环境五指令链上的成功率仅0.08%，展示长时序与跨环境泛化的难度；不是当前方法的上限。" }, { "section": "局限与复现条件", "note": "物体被简化为基础形状，属于仿真桌面场景；语言、颜色及域偏移会引起失败，不等于真实家务能力。" } ], "originalSha256": "c8ad4741782fa8fb773377fbf61b8772b86574f255599a229d3f6ad4d878c60e", "readingScopeSha256": "208fa969509dfe902cef9fbd1abda82eff156a48d7aa0d39de78826b7fe0792c", "textExtractionSha256": "cda2891193301e21199801c8daafb9f5202c64139cea93eea9ec58b1f5a256fd" }, "experimentNote": "区分单环境、多个已见环境和三环境训练/第四环境零样本测试；除单任务外还评估连续五条指令完成情况。", "contribution": "四套视觉布局中的34项操作技能组合成语言指令链，提供多模态观测和自由交互示教，用来测试语言落地与连续控制。", "whyUseful": "严格记录ABC→D等训练/测试组合、传感器配置、动作表示及指令链采样，避免把单技能成功率与整链成功率混合。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "MIT；仿真资产与依赖分别核对。", "codeReleaseScope": "已核验官方代码与文档；未运行复现" }, { "id": "arxiv-2109.11978", "title": "Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning", "titleZh": "通过大规模并行深度强化学习在数分钟内学会行走", "shortTitle": "Learning to Walk in Minutes", "date": "2021-09-24", "datePrecision": "day", "year": 2021, "url": "https://arxiv.org/abs/2109.11978", "codeUrl": "https://github.com/leggedrobotics/legged_gym", "category": "足式运动", "tags": [ "GPU并行", "课程学习", "Sim2Real" ], "tier": "classic", "summary": "把物理仿真、观测、回报、PPO及缓存全部留在GPU，用数千机器人同时采样。地形课程在同一网格上移动机器人切换难度，正确区分超时截断与失败终止；关节目标经学得的执行器模型转力矩，并施加噪声、摩擦和推动随机化。", "abstractZh": "把物理仿真、观测、回报、PPO及缓存全部留在GPU，用数千机器人同时采样。地形课程在同一网格上移动机器人切换难度，正确区分超时截断与失败终止；关节目标经学得的执行器模型转力矩，并施加噪声、摩擦和推动随机化。\n固定批量下扫描128到16384并行机器人，1500次更新、五次重复比较速度/回报。部署配置4096机器人、98304批量，在RTX A6000上不足20分钟训练；仿真还覆盖ANYmal B、带臂C、A1和Cassie，实机示范为ANYmal。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "ANYmal C", "ANYmal B（仿真）", "Unitree A1（仿真）", "Cassie（仿真）" ], "codeStatus": "open", "status": "官方legged_gym训练环境公开；依赖Isaac Gym与配套RL库。", "trainingNote": "GPU并行仿真、PPO、地形课程及动力学随机化。", "whyUseful": "应保留每环境采样长度、批量、超时bootstrap、地形课程和执行器模型，分别报告训练墙钟时间与实机安全速度。仿真A1/Cassie证据不能迁移成实机结论；未运行实现。", "contribution": "把物理仿真、观测、回报、PPO及缓存全部留在GPU，用数千机器人同时采样。地形课程在同一网格上移动机器人切换难度，正确区分超时截断与失败终止；关节目标经学得的执行器模型转力矩，并施加噪声、摩擦和推动随机化。", "limitations": "分钟级指特定GPU与固定任务的训练时间，不是无需模型、奖励调节或硬件标定。真实地形图和状态估计误差降低鲁棒性；不同形态仍需修改执行器、增益或奖励，不能称参数完全通用。", "caveats": "真实验证为ANYmal C；项目页其他机器人演示不能自动算作真机验证。", "evidence": [ { "url": "https://arxiv.org/abs/2109.11978", "note": "首发日期、并行训练、分钟级实验与真机迁移。" }, { "url": "https://leggedrobotics.github.io/legged_gym/", "note": "明确真实部署ANYmal C，区分额外机器人演示。" }, { "url": "https://github.com/leggedrobotics/legged_gym", "note": "作者官方训练环境及执行器网络、随机化和扰动组件。" }, { "url": "https://arxiv.org/pdf/2109.11978", "note": "§4.1 Figures 4–5：五次重复、并行/轨迹长度折中及仿真地形成功。" }, { "url": "https://arxiv.org/pdf/2109.11978", "note": "§4.2–4.3：A6000不足20分钟、其他形态为仿真、实机降速0.6m/s。" } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "运动控制", "强化学习" ], "robotFilters": [ "ANYmal C", "ANYmal B", "Unitree A1", "Cassie" ], "verificationNote": "2026-10-04核对论文、官方项目页及代码；只列已明确的真实机器人型号。", "original": { "id": "arxiv-2109.11978", "originalSourceUrl": "https://arxiv.org/pdf/2109.11978v3", "versionedOriginalUrl": "https://arxiv.org/pdf/2109.11978v3", "finalDownloadUrl": "https://arxiv.org/pdf/2109.11978", "sha256": "48014ee8af3e5a7fc26de2a5aac0de9ad593725075627c1a210d9d0f1e4ee62a", "bytes": 38462407, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:32.104150+00:00", "publicationDate": "2021/09/24", "publicationDateSourceUrl": "https://arxiv.org/abs/2109.11978", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2109.11978", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2109.11978", "pages": 14, "sourceIdentityStatus": "title_match", "sourceTitle": "Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning", "sourceVersion": "2109.11978v3", "sourceVersionDate": "2022/08/19", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.316275+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2109.11978" }, "originalAnalysis": { "id": "arxiv-2109.11978", "sourceUrl": "https://arxiv.org/pdf/2109.11978", "sectionsRead": [ "§2", "§3", "§4.1–4.3", "Figures 4–7", "§5" ], "methodsZh": "把物理仿真、观测、回报、PPO及缓存全部留在GPU，用数千机器人同时采样。地形课程在同一网格上移动机器人切换难度，正确区分超时截断与失败终止；关节目标经学得的执行器模型转力矩，并施加噪声、摩擦和推动随机化。", "experimentsZh": "固定批量下扫描128到16384并行机器人，1500次更新、五次重复比较速度/回报。部署配置4096机器人、98304批量，在RTX A6000上不足20分钟训练；仿真还覆盖ANYmal B、带臂C、A1和Cassie，实机示范为ANYmal。", "resultsZh": "作者发现2048–4096并行、约10万至20万批量取得较佳折中；并行过高导致每机器人轨迹过短而退化。仿真20厘米台阶接近全成功，25度以上上坡困难；实机能上下楼梯，但因高度图误差把最高线速度降到0.6米/秒。", "limitationsZh": "分钟级指特定GPU与固定任务的训练时间，不是无需模型、奖励调节或硬件标定。真实地形图和状态估计误差降低鲁棒性；不同形态仍需修改执行器、增益或奖励，不能称参数完全通用。", "reproductionZh": "应保留每环境采样长度、批量、超时bootstrap、地形课程和执行器模型，分别报告训练墙钟时间与实机安全速度。仿真A1/Cassie证据不能迁移成实机结论；未运行实现。", "experimentType": "both", "robots": [ "ANYmal C", "ANYmal B（仿真）", "Unitree A1（仿真）", "Cassie（仿真）" ], "evidenceNotes": [ { "section/page": "§4.1 Figures 4–5", "note": "五次重复、并行/轨迹长度折中及仿真地形成功。" }, { "section/page": "§4.2–4.3", "note": "A6000不足20分钟、其他形态为仿真、实机降速0.6m/s。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "固定批量下扫描128到16384并行机器人，1500次更新、五次重复比较速度/回报。部署配置4096机器人、98304批量，在RTX A6000上不足20分钟训练；仿真还覆盖ANYmal B、带臂C、A1和Cassie，实机示范为ANYmal。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2109.12098", "title": "CLIPort: What and Where Pathways for Robotic Manipulation", "titleZh": "CLIPort：机器人操作的语义与空间双通路", "shortTitle": "CLIPort", "date": "2021-09-24", "datePrecision": "day", "year": 2021, "url": "https://arxiv.org/abs/2109.12098", "codeUrl": "https://github.com/cliport/cliport", "category": "视觉语言操作", "tags": [ "语言条件", "CLIP", "抓放" ], "tier": "classic", "summary": "CLIPort融合冻结CLIP的语义通道与Transporter的空间等变通道，从俯视RGB-D和语言生成拾取/放置位置及离散旋转。以示范像素动作交叉熵训练，语义分支回答操作什么，空间分支保持精确几何；多任务按…", "abstractZh": "CLIPort融合冻结CLIP的语义通道与Transporter的空间等变通道，从俯视RGB-D和语言生成拾取/放置位置及离散旋转。以示范像素动作交叉熵训练，语义分支回答操作什么，空间分支保持精确几何；多任务按任务均匀采样。\n仿真Ravens/PyBullet含十语言任务，UR5e吸盘、三无噪声相机，训练1/10/100/1000示范并各测100实例，划分未见颜色/对象。真实Franka Panda九任务用179个图像动作对训练，每任务5–10测试回合。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "Franka Panda" ], "codeStatus": "open", "status": "官方训练与仿真评测代码公开。", "trainingNote": "使用预训练CLIP与任务示范训练语言条件策略。", "whyUseful": "复现需一致的俯视重建、动作旋转分箱、颜色/对象划分及验证选点流程；单任务20万更新、多任务60万，训练预算并不相同。真实Panda和仿真UR5e应分清，本轮未执行公开训练代码。", "contribution": "CLIPort融合冻结CLIP的语义通道与Transporter的空间等变通道，从俯视RGB-D和语言生成拾取/放置位置及离散旋转。以示范像素动作交叉熵训练，语义分支回答操作什么，空间分支保持精确几何；多任务按任务均匀采样。", "limitations": "指标采用Ravens部分完成得分，百分比不能都解释成完整回合成功率；环境用oracle判断结束。两阶段拾放原语无法直接覆盖完整6DoF、部分可观测长任务或多指连续控制，真实测试样本有限。", "caveats": "真实平台为Franka Panda；不能从Ravens仿真资产反推真实硬件。", "evidence": [ { "url": "https://arxiv.org/abs/2109.12098", "note": "首发日期与摘要。" }, { "url": "https://arxiv.org/pdf/2109.12098", "note": "正文和附录D明确真实Franka Panda及并联夹爪。" }, { "url": "https://cliport.github.io/", "note": "作者项目页提供官方代码链接与多任务设置。" }, { "url": "https://github.com/cliport/cliport", "note": "已打开验证官方仓库。" }, { "url": "https://arxiv.org/pdf/2109.12098", "note": "§4.1：部分完成评分和oracle终止；UR5e为仿真。" }, { "url": "https://arxiv.org/pdf/2109.12098", "note": "Table 2 / §4.3：真实Panda九任务、179图像动作样本、5–10测试。" }, { "url": "https://arxiv.org/pdf/2109.12098", "note": "§5：受两步原语、部分可观测和6DoF控制限制。" } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "操作与抓取" ], "robotFilters": [ "Franka Panda" ], "verificationNote": "2026-10-04核对论文与官方项目/代码；未运行复现。", "original": { "id": "arxiv-2109.12098", "originalSourceUrl": "https://arxiv.org/pdf/2109.12098v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2109.12098v1", "finalDownloadUrl": "https://arxiv.org/pdf/2109.12098", "sha256": "3cc19664beabda203cb1a1ad86659f75f60a099189820d92d59d20067c9ab003", "bytes": 8143919, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:27.363965+00:00", "publicationDate": "2021/09/24", "publicationDateSourceUrl": "https://arxiv.org/abs/2109.12098", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2109.12098", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2109.12098", "pages": 24, "sourceIdentityStatus": "title_match", "sourceTitle": "CLIPort: What and Where Pathways for Robotic Manipulation", "sourceVersion": "2109.12098v1", "sourceVersionDate": "2021/09/24", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.279694+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2109.12098" }, "originalAnalysis": { "id": "arxiv-2109.12098", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090144+00:00", "sourceUrl": "https://arxiv.org/pdf/2109.12098", "sectionsRead": [ "§3", "§4.1–4.3", "Tables 1–2", "§5", "Appendix C–E" ], "methodsZh": "CLIPort融合冻结CLIP的语义通道与Transporter的空间等变通道，从俯视RGB-D和语言生成拾取/放置位置及离散旋转。以示范像素动作交叉熵训练，语义分支回答操作什么，空间分支保持精确几何；多任务按任务均匀采样。", "experimentsZh": "仿真Ravens/PyBullet含十语言任务，UR5e吸盘、三无噪声相机，训练1/10/100/1000示范并各测100实例，划分未见颜色/对象。真实Franka Panda九任务用179个图像动作对训练，每任务5–10测试回合。", "resultsZh": "作者真实实验分任务得分约55–75%，简单积木约70%；语义与空间融合及跨任务数据有助于未见属性。实机少量示范也暴露语言与对象颜色共现偏差，不能仅凭语义预训练宣称正确理解所有指令。", "limitationsZh": "指标采用Ravens部分完成得分，百分比不能都解释成完整回合成功率；环境用oracle判断结束。两阶段拾放原语无法直接覆盖完整6DoF、部分可观测长任务或多指连续控制，真实测试样本有限。", "reproductionZh": "复现需一致的俯视重建、动作旋转分箱、颜色/对象划分及验证选点流程；单任务20万更新、多任务60万，训练预算并不相同。真实Panda和仿真UR5e应分清，本轮未执行公开训练代码。", "experimentType": "both", "robots": [ "Franka Panda" ], "evidenceNotes": [ { "section/page": "§4.1", "note": "部分完成评分和oracle终止；UR5e为仿真。" }, { "section/page": "Table 2 / §4.3", "note": "真实Panda九任务、179图像动作样本、5–10测试。" }, { "section/page": "§5", "note": "受两步原语、部分可观测和6DoF控制限制。" } ], "corrections": { "metricNote": "Ravens得分可含部分完成，不能统一称完整任务成功率。" } }, "experimentNote": "仿真Ravens/PyBullet含十语言任务，UR5e吸盘、三无噪声相机，训练1/10/100/1000示范并各测100实例，划分未见颜色/对象。真实Franka Panda九任务用179个图像动作对训练，每任务5–10测试回合。", "analysisVerifiedAt": "2026-10-04T13:50:35.090144+00:00" }, { "id": "arxiv-2108.10869", "title": "DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras", "date": "2021-08-24", "url": "https://arxiv.org/abs/2108.10869v2", "titleZh": "DROID-SLAM：面向单目、双目与RGB-D相机的深度视觉SLAM", "abstractZh": "循环网络预测稠密对应修正及置信度，通过可微稠密束调整联合更新相机位姿和逐像素逆深度；单目训练后可接入双目或深度约束。\n在合成TartanAir训练，评估TartanAir、EuRoC、TUM-RGB-D和ETH3D等视频序列；属于定位建图数据评测，不是机器人任务闭环。", "summary": "循环网络预测稠密对应修正及置信度，通过可微稠密束调整联合更新相机位姿和逐像素逆深度；单目训练后可接入双目或深度约束。", "experimentType": "data", "robots": [], "tags": [ "导航与建图" ], "limitations": "主要瓶颈是显存与计算：部分长序列需要24GB显存，TartanAir约8fps；不能将所有配置称为实时。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：代表性的学习式视觉SLAM研究，连接稠密对应预测与几何优化。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2108.10869v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2108.10869v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/princeton-vl/DROID-SLAM", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/princeton-vl/DROID-SLAM", "category": "导航与建图", "year": 2021, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "train.py 已发布，训练使用 TartanAir。 提供数据下载/整理入口；数据归各自数据集许可。 官方 droid.pth 下载已列出。 相机算法不等于任意整机已集成；需要正确标定，异步多 GPU 可视化存在已知限制。 本次未运行训练、复现实验或实机控制。", "directions": [ "导航与建图" ], "robotFilters": [], "original": { "id": "arxiv-2108.10869", "originalSourceUrl": "https://arxiv.org/pdf/2108.10869v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2108.10869v2", "finalDownloadUrl": "https://arxiv.org/pdf/2108.10869v2", "sha256": "0e15f20bff5f7ceac468c2899baafc8c114a3018c7a7b213eed7c402e5ccec90", "bytes": 11003883, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:31.058980+00:00", "publicationDate": "2021/08/24", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2108.10869v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2108.10869v2", "pages": 15, "sourceIdentityStatus": "title_match", "sourceTitle": "DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras", "sourceVersion": "2108.10869v2", "sourceVersionDate": "2022/02/02", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2108.10869v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2108.10869" }, "originalAnalysis": { "id": "arxiv-2108.10869", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2108.10869v2", "sourceVersion": "2108.10869v2", "sectionsRead": [ "PDF第7页选段（提取文本L402–419）", "PDF第9–10页选段（提取文本L994–1011）", "PDF第3页选段（提取文本L150–168）", "PDF第5页选段（提取文本L265–310）", "PDF第9页选段（提取文本L974–990）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "循环网络预测稠密对应修正及置信度，通过可微稠密束调整联合更新相机位姿和逐像素逆深度；单目训练后可接入双目或深度约束。", "experimentsZh": "在合成TartanAir训练，评估TartanAir、EuRoC、TUM-RGB-D和ETH3D等视频序列；属于定位建图数据评测，不是机器人任务闭环。", "resultsZh": "跨数据集结果较当时基线减少跟踪失败；论文在ETH3D测试中跟踪30/32条RGB-D序列，同时说明无图像暗场序列未提交。", "limitationsZh": "主要瓶颈是显存与计算：部分长序列需要24GB显存，TartanAir约8fps；不能将所有配置称为实时。", "reproductionZh": "保留图像降采样、跳帧、前后端GPU配置、相机标定和轨迹对齐协议，报告失败率而非仅成功序列误差。", "experimentType": "data", "robots": [], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "循环网络预测稠密对应修正及置信度，通过可微稠密束调整联合更新相机位姿和逐像素逆深度；单目训练后可接入双目或深度约束。" }, { "section": "原文实验与结果相关选段", "note": "在合成TartanAir训练，评估TartanAir、EuRoC、TUM-RGB-D和ETH3D等视频序列；属于定位建图数据评测，不是机器人任务闭环。 跨数据集结果较当时基线减少跟踪失败；论文在ETH3D测试中跟踪30/32条RGB-D序列，同时说明无图像暗场序列未提交。" }, { "section": "局限与复现条件", "note": "主要瓶颈是显存与计算：部分长序列需要24GB显存，TartanAir约8fps；不能将所有配置称为实时。" } ], "originalSha256": "0e15f20bff5f7ceac468c2899baafc8c114a3018c7a7b213eed7c402e5ccec90", "readingScopeSha256": "637ea8a150e3d34e8c718e14c03f9b19e93d47031d4eda9e1d1161172ce20a9c", "textExtractionSha256": "ad25c8f04e3c82494ca2059b2d0a6520855a108e01f7c9f9821cc0cd80e1ec29" }, "experimentNote": "在合成TartanAir训练，评估TartanAir、EuRoC、TUM-RGB-D和ETH3D等视频序列；属于定位建图数据评测，不是机器人任务闭环。", "contribution": "循环网络预测稠密对应修正及置信度，通过可微稠密束调整联合更新相机位姿和逐像素逆深度；单目训练后可接入双目或深度约束。", "whyUseful": "保留图像降采样、跳帧、前后端GPU配置、相机标定和轨迹对齐协议，报告失败率而非仅成功序列误差。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "主代码 BSD-3-Clause；TartanAir 等数据集和第三方工具分别核对。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-2108.05877", "title": "DexMV: Imitation Learning for Dexterous Manipulation from Human Videos", "titleZh": "DexMV：从人类视频学习灵巧操作", "shortTitle": "DexMV", "date": "2021-08-12", "datePrecision": "day", "year": 2021, "url": "https://arxiv.org/abs/2108.05877", "paperUrl": "https://arxiv.org/abs/2108.05877", "codeUrl": "https://github.com/yzqin/dexmv-learn", "summary": "DexMV先从真实RGB-D视频估计MANO人手及物体六维姿态，再用保留手内相对结构的优化重定向为机器人关节轨迹。最小jerk连续拟合后用解析逆动力学估计动作，将30Hz视频对齐120Hz仿真，供DAPG等示范…", "abstractZh": "DexMV先从真实RGB-D视频估计MANO人手及物体六维姿态，再用保留手内相对结构的优化重定向为机器人关节轨迹。最小jerk连续拟合后用解析逆动力学估计动作，将30Hz视频对齐120Hz仿真，供DAPG等示范增强强化学习使用。\nAdroit多指手仿真中评估五物体Relocate、倒粒子及香蕉放入杯子；比较TRPO、SOIL、GAIL+、DAPG，三随机种子，Relocate每种子100次。另测试单/双相机、重定向项、示范数量、物体形状尺寸及新类别。", "category": "灵巧手 / 人类视频示范", "tags": [ "视频示范", "动作重定向", "示范增强RL", "数据集" ], "directions": [ "灵巧手", "操作与抓取", "数据集与基准", "模仿学习" ], "tier": "classic", "experimentType": "sim", "robots": [ "Adroit Robotic Hand（仿真）" ], "robotNote": "Adroit仅在MuJoCo中使用；真实内容是人手示范视频。", "codeStatus": "open", "status": "作者学习代码与配套仿真仓库公开。", "trainingNote": "先准备视频恢复与重定向示范，再运行示范增强策略学习；学习仓库需配套dexmv-sim和数据。", "contribution": "DexMV先从真实RGB-D视频估计MANO人手及物体六维姿态，再用保留手内相对结构的优化重定向为机器人关节轨迹。最小jerk连续拟合后用解析逆动力学估计动作，将30Hz视频对齐120Hz仿真，供DAPG等示范增强强化学习使用。", "whyUseful": "先检验手物坐标、时序平滑与逆动力学单位，再固定TRPO预算公平比较状态-only和状态—动作示范。务必区分颗粒比例、Inside Score与二元成功率；先从Relocate而非流体任务建立可复现基线。", "limitations": "真实的是人类示范视频，策略执行全部在模拟手中，不能算灵巧手实机验证。姿态估计受遮挡与相机数影响；动作由模型推导而非真实接触力测量，仿真动力学和物体几何误差可能阻碍迁移。", "evidence": [ { "url": "https://arxiv.org/abs/2108.05877", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/pdf/2108.05877", "note": "正文明确MuJoCo、Adroit模型和三类仿真任务。" }, { "url": "https://github.com/yzqin/dexmv-learn", "note": "作者仓库说明包含策略训练/推理代码并链接仿真与示范数据。" }, { "url": "https://yzqin.github.io/dexmv/", "note": "作者项目展示示范转换及对象变化评估。" }, { "url": "https://arxiv.org/pdf/2108.05877", "note": "5–7：RGB-D姿态、最小jerk拟合、解析动作与30/120Hz对齐。" }, { "url": "https://arxiv.org/pdf/2108.05877", "note": "8.1–8.3：Adroit任务、三种子、27.2%粒子指标及算法差异。" }, { "url": "https://arxiv.org/pdf/2108.05877v5", "note": "5–7：RGB-D姿态、最小jerk拟合、解析动作与30/120Hz对齐。" }, { "url": "https://arxiv.org/pdf/2108.05877v5", "note": "8.1–8.3：Adroit任务、三种子、27.2%粒子指标及算法差异。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2021：把人类视频转为灵巧操作学习信号，连接视觉手物理解与机器人训练。", "projectUrl": null, "robotFilters": [], "original": { "id": "arxiv-2108.05877", "originalSourceUrl": "https://arxiv.org/pdf/2108.05877v5", "versionedOriginalUrl": "https://arxiv.org/pdf/2108.05877v5", "finalDownloadUrl": "https://arxiv.org/pdf/2108.05877", "sha256": "9b2e0e5207253be4594aaf3dfca6db8d38527e937738239409e5536ef3cf57c4", "bytes": 8540406, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:35.269164+00:00", "publicationDate": "2021/08/12", "publicationDateSourceUrl": "https://arxiv.org/abs/2108.05877", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2108.05877", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2108.05877", "pages": 25, "sourceIdentityStatus": "title_match", "sourceTitle": "DexMV: Imitation Learning for Dexterous Manipulation from Human Videos", "sourceVersion": "2108.05877v5", "sourceVersionDate": "2022/07/06", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.324399+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2108.05877" }, "originalAnalysis": { "id": "arxiv-2108.05877", "sourceUrl": "https://arxiv.org/pdf/2108.05877v5", "sectionsRead": [ "3–7", "8.1–8.3", "9", "Table 1" ], "methodsZh": "DexMV先从真实RGB-D视频估计MANO人手及物体六维姿态，再用保留手内相对结构的优化重定向为机器人关节轨迹。最小jerk连续拟合后用解析逆动力学估计动作，将30Hz视频对齐120Hz仿真，供DAPG等示范增强强化学习使用。", "experimentsZh": "Adroit多指手仿真中评估五物体Relocate、倒粒子及香蕉放入杯子；比较TRPO、SOIL、GAIL+、DAPG，三随机种子，Relocate每种子100次。另测试单/双相机、重定向项、示范数量、物体形状尺寸及新类别。", "resultsZh": "示范增强方法在五类移动任务均优于纯RL，但最佳算法随对象变化，糖盒上SOIL更好。倒水DAPG平均仅27.2%粒子进入目标；放入任务按体积覆盖及成功率评估，说明视频示范提供帮助却未解决全部精细操作。", "limitationsZh": "真实的是人类示范视频，策略执行全部在模拟手中，不能算灵巧手实机验证。姿态估计受遮挡与相机数影响；动作由模型推导而非真实接触力测量，仿真动力学和物体几何误差可能阻碍迁移。", "reproductionZh": "先检验手物坐标、时序平滑与逆动力学单位，再固定TRPO预算公平比较状态-only和状态—动作示范。务必区分颗粒比例、Inside Score与二元成功率；先从Relocate而非流体任务建立可复现基线。", "experimentType": "sim", "robots": [ "Adroit Robotic Hand（仿真）" ], "corrections": [ "人类真实视频采集不等于机器人实机实验；DexMV本篇控制评估为仿真。" ], "evidenceNotes": [ { "note": "RGB-D姿态、最小jerk拟合、解析动作与30/120Hz对齐。", "section": "5–7" }, { "note": "Adroit任务、三种子、27.2%粒子指标及算法差异。", "section": "8.1–8.3" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2108.05877v5", "originalSha256": "9b2e0e5207253be4594aaf3dfca6db8d38527e937738239409e5536ef3cf57c4" }, "experimentNote": "Adroit多指手仿真中评估五物体Relocate、倒粒子及香蕉放入杯子；比较TRPO、SOIL、GAIL+、DAPG，三随机种子，Relocate每种子100次。另测试单/双相机、重定向项、示范数量、物体形状尺寸及新类别。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2108.03298", "title": "What Matters in Learning from Offline Human Demonstrations for Robot Manipulation", "titleZh": "从离线人类示范学习机器人操作：哪些因素最重要", "shortTitle": "robomimic", "date": "2021-08-06", "year": 2021, "url": "https://arxiv.org/abs/2108.03298", "codeUrl": "https://github.com/ARISE-Initiative/robomimic", "category": "数据集与基准", "tags": [ "模仿学习", "离线强化学习", "可复现", "数据质量" ], "tier": "classic", "summary": "这是一项离线人类示范学习的系统比较，评估BC、带历史的BC-RNN、分层HBC、BCQ、CQL与IRIS，而非单一新策略。数据分机器生成、熟练单人200条、六位不同熟练度操作者共300条；控制数据质量、历史、观…", "abstractZh": "这是一项离线人类示范学习的系统比较，评估BC、带历史的BC-RNN、分层HBC、BCQ、CQL与IRIS，而非单一新策略。数据分机器生成、熟练单人200条、六位不同熟练度操作者共300条；控制数据质量、历史、观察模态、网络结构和检查点选择来分析性能来源。\n五项robosuite仿真任务覆盖抬升、分拣、套螺母、双臂传递与工具悬挂，三项在Franka Emika Panda实机验证。仿真每检查点50次在线展开、三种子，并报告训练期间最好结果；实机每任务200条示范，只评最终检查点30次，二者选模协议不同。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "Franka Emika Panda" ], "codeStatus": "open", "status": "MIT 框架代码、研究数据和训练模型公开。", "trainingNote": "当前仓库新增多种后续算法；复现 2021 研究应使用对应版本、数据划分与配置。", "whyUseful": "复现应固定数据版本、人类子集、回放序列、GMM动作头、ResNet/腕视角和训练预算；低维与图像设置不能混算。报告最终策略及选模成本，并保留工具悬挂低成功结果，公开框架本身不代表用户硬件可直接运行。", "contribution": "这是一项离线人类示范学习的系统比较，评估BC、带历史的BC-RNN、分层HBC、BCQ、CQL与IRIS，而非单一新策略。数据分机器生成、熟练单人200条、六位不同熟练度操作者共300条；控制数据质量、历史、观察模态、网络结构和检查点选择来分析性能来源。", "limitations": "结论限于当时算法和选定数据，不能推导离线RL普遍无效。仿真挑最好在线检查点带有不可直接用于真实离线部署的优势；验证损失或最后检查点的性能可显著下降。更大或更杂的数据不必然更好，极精细长程实机任务仍很弱。", "caveats": "robomimic 是框架名称，本条对应其 2021 年研究论文，非后续新增算法。", "evidence": [ { "url": "https://arxiv.org/abs/2108.03298", "note": "首发日期、六算法与仿真/实机任务数量。" }, { "url": "https://arxiv.org/html/2108.03298v1", "note": "附录硬件明确仿真与实机均使用 Panda。" }, { "url": "https://github.com/ARISE-Initiative/robomimic", "note": "MIT、数据入口和版本更新。" }, { "url": "https://arxiv.org/pdf/2108.03298v2", "note": "3.2/3.3：PH200 demonstrations;MH6×50;3 seeds;50 online rollouts/checkpoint;report maximum during training." }, { "url": "https://arxiv.org/pdf/2108.03298v2", "note": "4.4/4.5：GMM/history/observation design matter;offline validation-loss model selection may fail." }, { "url": "https://arxiv.org/pdf/2108.03298v2", "note": "4.7：Panda real datasets200 demos/task;30 rollouts final checkpoint;96.7/73.3/3.3%;no real hyperparameter tuning." }, { "url": "https://arxiv.org/pdf/2108.03298v2", "note": "附录E：Franka Emika Panda used in simulation and real workspaces." } ], "verification": "verified", "verifiedAt": "2026-10-04", "fullTextTranslation": "未提供", "directions": [ "操作与抓取", "数据集与基准", "强化学习", "模仿学习" ], "robotFilters": [ "Franka Panda" ], "verificationNote": "已核论文及官方仓库；与后续软件功能明确区分。", "original": { "id": "arxiv-2108.03298", "originalSourceUrl": "https://arxiv.org/pdf/2108.03298v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2108.03298v2", "finalDownloadUrl": "https://arxiv.org/pdf/2108.03298", "sha256": "18283ff9cbb9645de4ccf32ca2de472f452e89761287e7ed578e3a1066bea5f1", "bytes": 10782355, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:39.856810+00:00", "publicationDate": "2021/08/06", "publicationDateSourceUrl": "https://arxiv.org/abs/2108.03298", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2108.03298", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2108.03298", "pages": 36, "sourceIdentityStatus": "title_match", "sourceTitle": "What Matters in Learning from Offline Human Demonstrations for Robot Manipulation", "sourceVersion": "2108.03298v2", "sourceVersionDate": "2021/09/25", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.332689+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2108.03298" }, "originalAnalysis": { "id": "arxiv-2108.03298", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739405+00:00", "sourceUrl": "https://arxiv.org/pdf/2108.03298v2", "sectionsRead": [ "3.1–3.3 任务、数据来源与评估", "4.1–4.7 六算法、观察/超参/模型选择消融及实机", "附录E 硬件" ], "methodsZh": "这是一项离线人类示范学习的系统比较，评估BC、带历史的BC-RNN、分层HBC、BCQ、CQL与IRIS，而非单一新策略。数据分机器生成、熟练单人200条、六位不同熟练度操作者共300条；控制数据质量、历史、观察模态、网络结构和检查点选择来分析性能来源。", "experimentsZh": "五项robosuite仿真任务覆盖抬升、分拣、套螺母、双臂传递与工具悬挂，三项在Franka Emika Panda实机验证。仿真每检查点50次在线展开、三种子，并报告训练期间最好结果；实机每任务200条示范，只评最终检查点30次，二者选模协议不同。", "resultsZh": "作者发现历史信息对长程与多人数据尤重要，BC-RNN通常强于BC；BCQ/CQL在混合人类数据上明显弱于其机器数据表现。实机Lift、Can、Tool Hang为96.7%、73.3%、3.3%；去图像随机移位或腕相机使Can降至26.7%、43.3%，显示观察工程的重要性。", "limitationsZh": "结论限于当时算法和选定数据，不能推导离线RL普遍无效。仿真挑最好在线检查点带有不可直接用于真实离线部署的优势；验证损失或最后检查点的性能可显著下降。更大或更杂的数据不必然更好，极精细长程实机任务仍很弱。", "reproductionZh": "复现应固定数据版本、人类子集、回放序列、GMM动作头、ResNet/腕视角和训练预算；低维与图像设置不能混算。报告最终策略及选模成本，并保留工具悬挂低成功结果，公开框架本身不代表用户硬件可直接运行。", "evidenceNotes": [ { "section/page": "3.2/3.3", "note": "PH200 demonstrations;MH6×50;3 seeds;50 online rollouts/checkpoint;report maximum during training." }, { "section/page": "4.4/4.5", "note": "GMM/history/observation design matter;offline validation-loss model selection may fail." }, { "section/page": "4.7", "note": "Panda real datasets200 demos/task;30 rollouts final checkpoint;96.7/73.3/3.3%;no real hyperparameter tuning." }, { "section/page": "附录E", "note": "Franka Emika Panda used in simulation and real workspaces." } ], "experimentType": "both", "robots": [ "Franka Emika Panda" ], "corrections": [ "区分仿真最佳在线选模与实机最终检查点，避免把仿真表现当严格离线选模结果。" ], "sourceVersion": "2108.03298v2", "originalSha256": "18283ff9cbb9645de4ccf32ca2de472f452e89761287e7ed578e3a1066bea5f1" }, "experimentNote": "五项robosuite仿真任务覆盖抬升、分拣、套螺母、双臂传递与工具悬挂，三项在Franka Emika Panda实机验证。仿真每检查点50次在线展开、三种子，并报告训练期间最好结果；实机每任务200条示范，只评最终检查点30次，二者选模协议不同。", "analysisVerifiedAt": "2026-10-04T13:47:58.739405+00:00" }, { "id": "arxiv-2107.06829", "title": "FAST-LIO2: Fast Direct LiDAR-inertial Odometry", "titleZh": "FAST-LIO2：快速直接激光惯性里程计", "date": "2021-07-14", "datePrecision": "day", "year": 2021, "url": "https://arxiv.org/abs/2107.06829", "paperUrl": "https://arxiv.org/abs/2107.06829", "codeUrl": "https://github.com/hku-mars/FAST_LIO", "summary": "FAST-LIO2直接将原始LiDAR点配准到地图，结合IMU反向去畸变和流形迭代卡尔曼滤波。增量ikd-Tree支持点插入、删除、降采样和重平衡，避免每帧重建索引，使更密集测量与高频定位能够同时使用。", "abstractZh": "FAST-LIO2直接将原始LiDAR点配准到地图，结合IMU反向去畸变和流形迭代卡尔曼滤波。增量ikd-Tree支持点插入、删除、降采样和重平衡，避免每帧重建索引，使更密集测量与高频定位能够同时使用。\n数据结构与里程计分别评估，多数据集19序列按有效真值计算RMSE或闭环端点漂移。公平比较时关闭LILI-OM/LIO-SAM回环；另在手持设备、定制小型四旋翼翻转及较大型航测机上使用Livox Avia和机载计算测试。", "category": "激光惯性 / 直接配准", "tags": [ "FAST-LIO2", "迭代卡尔曼滤波", "ikd-Tree", "LiDAR-IMU" ], "directions": [ "导航与建图" ], "tier": "classic", "experimentType": "real", "robots": [ "定制280毫米轴距四旋翼", "750毫米轴距航测四旋翼" ], "robotFilters": [], "codeStatus": "open", "status": "实现已开源", "trainingNote": "无需神经网络训练；官方 FAST_LIO 仓库含 FAST-LIO2 实现、配置和样例数据。主仓库为 GPL-2.0，ikd-Tree 等依赖需另行核对。", "contribution": "FAST-LIO2直接将原始LiDAR点配准到地图，结合IMU反向去畸变和流形迭代卡尔曼滤波。增量ikd-Tree支持点插入、删除、降采样和重平衡，避免每帧重建索引，使更密集测量与高频定位能够同时使用。", "whyUseful": "复现应记录扫描率、点数、地图边长、LiDAR-IMU标定和时间同步，按论文关闭基线回环才比较同范围。无需神经训练，公开FAST_LIO与ikd-Tree为主要入口；本轮没有重跑序列或飞行。", "limitations": "系统是无回环修正的里程计，长期仍会漂移；较大地图可能引入错误旧点匹配。部分航测GPS轨迹仅视觉核对，未提供定量真值；成功快速运动例不能替代退化几何、动态场景的全面可靠性测试。", "license": "GPL-2.0（FAST_LIO 主仓库）", "evidence": [ { "url": "https://arxiv.org/abs/2107.06829", "note": "首版日期、直接注册、ikd-Tree、19 段基准及真实平台测试。" }, { "url": "https://github.com/hku-mars/FAST_LIO", "note": "官方 FAST-LIO2 发布入口，仓库许可证显示 GPL-2.0；不能沿用第三方索引的 MIT 标记。" }, { "url": "https://arxiv.org/pdf/2107.06829", "note": "§VI-C：19序列，基线回环关闭；该系统本身无回环。" }, { "url": "https://arxiv.org/pdf/2107.06829", "note": "§VII-B / Fig.11：翻转1198deg/s峰值，2.01ms均时。" }, { "url": "https://arxiv.org/pdf/2107.06829", "note": "§VII-C：航测约20–24ms，GPS量化轨迹不可用。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "以直接配准和增量空间索引提升激光惯性里程计的实时性与扫描模式适应性。", "experimentNote": "数据结构与里程计分别评估，多数据集19序列按有效真值计算RMSE或闭环端点漂移。公平比较时关闭LILI-OM/LIO-SAM回环；另在手持设备、定制小型四旋翼翻转及较大型航测机上使用Livox Avia和机载计算测试。", "projectUrl": null, "shortTitle": "", "original": { "id": "arxiv-2107.06829", "originalSourceUrl": "https://arxiv.org/pdf/2107.06829v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2107.06829v1", "finalDownloadUrl": "https://arxiv.org/pdf/2107.06829", "sha256": "ba5072e06e75dec5222bcb528143f3aa3238d6f6692bac1e22a9fe1bc55a15dd", "bytes": 13481318, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:45.184925+00:00", "publicationDate": "2021/07/14", "publicationDateSourceUrl": "https://arxiv.org/abs/2107.06829", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2107.06829", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2107.06829", "pages": 19, "sourceIdentityStatus": "title_match", "sourceTitle": "FAST-LIO2: Fast Direct LiDAR-inertial Odometry", "sourceVersion": "2107.06829v1", "sourceVersionDate": "2021/07/14", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.342868+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2107.06829" }, "originalAnalysis": { "id": "arxiv-2107.06829", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090145+00:00", "sourceUrl": "https://arxiv.org/pdf/2107.06829", "sectionsRead": [ "§III–V", "§VI", "Tables IV–VI", "§VII", "Figs.9–14", "§VIII" ], "methodsZh": "FAST-LIO2直接将原始LiDAR点配准到地图，结合IMU反向去畸变和流形迭代卡尔曼滤波。增量ikd-Tree支持点插入、删除、降采样和重平衡，避免每帧重建索引，使更密集测量与高频定位能够同时使用。", "experimentsZh": "数据结构与里程计分别评估，多数据集19序列按有效真值计算RMSE或闭环端点漂移。公平比较时关闭LILI-OM/LIO-SAM回环；另在手持设备、定制小型四旋翼翻转及较大型航测机上使用Livox Avia和机载计算测试。", "resultsZh": "作者报告方法或其变体在19序列中18个表现最佳；翻转平均/峰值角速912/1198度每秒，单扫描平均2.01毫秒。航测大场景平均处理约19.6–23.9毫秒，因此100Hz并不是所有地图/配置都保证。", "limitationsZh": "系统是无回环修正的里程计，长期仍会漂移；较大地图可能引入错误旧点匹配。部分航测GPS轨迹仅视觉核对，未提供定量真值；成功快速运动例不能替代退化几何、动态场景的全面可靠性测试。", "reproductionZh": "复现应记录扫描率、点数、地图边长、LiDAR-IMU标定和时间同步，按论文关闭基线回环才比较同范围。无需神经训练，公开FAST_LIO与ikd-Tree为主要入口；本轮没有重跑序列或飞行。", "experimentType": "real", "robots": [ "定制280毫米轴距四旋翼", "750毫米轴距航测四旋翼" ], "evidenceNotes": [ { "section/page": "§VI-C", "note": "19序列，基线回环关闭；该系统本身无回环。" }, { "section/page": "§VII-B / Fig.11", "note": "翻转1198deg/s峰值，2.01ms均时。" }, { "section/page": "§VII-C", "note": "航测约20–24ms，GPS量化轨迹不可用。" } ] }, "analysisVerifiedAt": "2026-10-04T13:50:35.090145+00:00" }, { "id": "arxiv-2107.04034", "title": "RMA: Rapid Motor Adaptation for Legged Robots", "titleZh": "RMA：足式机器人的快速运动适应", "shortTitle": "RMA", "date": "2021-07-08", "datePrecision": "day", "year": 2021, "url": "https://arxiv.org/abs/2107.04034", "codeUrl": "https://github.com/antonilo/rl_locomotion", "category": "足式运动", "tags": [ "在线适应", "Sim2Real", "四足" ], "tier": "classic", "summary": "先用特权质量、摩擦和电机等参数编码的8维环境潜变量训练基础行走策略，再学习50步历史到潜变量的适应网络。适应器在自身造成的轨迹上监督训练；部署只需本体传感，基础策略100Hz、适应器10Hz异步运行，不进行在线…", "abstractZh": "先用特权质量、摩擦和电机等参数编码的8维环境潜变量训练基础行走策略，再学习50步历史到潜变量的适应网络。适应器在自身造成的轨迹上监督训练；部署只需本体传感，基础策略100Hz、适应器10Hz异步运行，不进行在线梯度微调。\nRaiSim中改变摩擦、载荷、重心、电机能力和地形，测试范围更宽且回合内变化更频繁；三训练种子各1000回合。真实全部采用Unitree A1，比较原控制器、无适应和RMA，室内多数条件五次，严重失败时提前停在两次。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "Unitree A1" ], "codeStatus": "open", "status": "项目页链接后续RMA衍生实现；并非原论文完整代码原样发布。", "trainingNote": "仿真训练基础策略与适应模块；实机在线更新隐含状态。", "whyUseful": "需复现环境编码器、on-policy适应数据、奖励课程与随机化范围，严格区分潜变量更新和权重更新。真实评测应披露所有尝试及安全早停，以免夸大成功；未操作硬件。", "contribution": "先用特权质量、摩擦和电机等参数编码的8维环境潜变量训练基础行走策略，再学习50步历史到潜变量的适应网络。适应器在自身造成的轨迹上监督训练；部署只需本体传感，基础策略100Hz、适应器10Hz异步运行，不进行在线梯度微调。", "limitations": "无外部感知难提前应对大落差及多腿遮挡，仍发生跌倒。实机重复少且部分方法早停；图3图内与图注个别成功率不一致，本分析不把这些数字当作确定结果。训练分布需覆盖重要实际变化。", "caveats": "代码链接现为Cross-Modal Supervision的衍生训练库，复现时须检查与RMA原文差异。", "evidence": [ { "url": "https://roboticsproceedings.org/rss17/p011.pdf", "note": "RSS论文第III-B节以状态/动作历史估计隐含条件，第IV节明确所有真实实验使用Unitree A1。" }, { "url": "https://arxiv.org/abs/2107.04034", "note": "首发日期、双模块方法及A1无微调部署。" }, { "url": "https://ashish-kmr.github.io/rma-legged-robots/", "note": "作者项目页与现有Code链接。" }, { "url": "https://github.com/antonilo/rl_locomotion", "note": "README明确基于RMA的后续CMS训练代码，GPL-3.0许可。" }, { "url": "https://arxiv.org/pdf/2107.04034", "note": "§III–IV：50步历史、8维潜变量、100/10Hz异步与A1硬件。" }, { "url": "https://arxiv.org/pdf/2107.04034", "note": "Table II; §V; Figure 3：三种子千回合、实机早停，以及图内/图注不一致。" } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "运动控制" ], "robotFilters": [ "Unitree A1" ], "verificationNote": "2026-10-04核对论文与项目当前代码指向；原版与衍生实现已区分。", "original": { "id": "arxiv-2107.04034", "originalSourceUrl": "https://arxiv.org/pdf/2107.04034v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2107.04034v1", "finalDownloadUrl": "https://arxiv.org/pdf/2107.04034", "sha256": "d8660da881851756287efab24d85f5c629ba97878f70297eb429a8e8a9a57846", "bytes": 16534863, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:45.178010+00:00", "publicationDate": "2021/07/08", "publicationDateSourceUrl": "https://arxiv.org/abs/2107.04034", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2107.04034", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2107.04034", "pages": 15, "sourceIdentityStatus": "title_match", "sourceTitle": "RMA: Rapid Motor Adaptation for Legged Robots", "sourceVersion": "2107.04034v1", "sourceVersionDate": "2021/07/08", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.355542+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2107.04034" }, "originalAnalysis": { "id": "arxiv-2107.04034", "sourceUrl": "https://arxiv.org/pdf/2107.04034", "sectionsRead": [ "§III-A–C", "§IV", "§V", "Table II", "§VI" ], "methodsZh": "先用特权质量、摩擦和电机等参数编码的8维环境潜变量训练基础行走策略，再学习50步历史到潜变量的适应网络。适应器在自身造成的轨迹上监督训练；部署只需本体传感，基础策略100Hz、适应器10Hz异步运行，不进行在线梯度微调。", "experimentsZh": "RaiSim中改变摩擦、载荷、重心、电机能力和地形，测试范围更宽且回合内变化更频繁；三训练种子各1000回合。真实全部采用Unitree A1，比较原控制器、无适应和RMA，室内多数条件五次，严重失败时提前停在两次。", "resultsZh": "作者报告仿真RMA成功73.5%，无适应52.1%、特权专家76.2%；真实展示沙地、植被、泡沫、油面和额外载荷，油面报告90%。潜变量随滑移改变支持快速适应解释，但不是对真实摩擦等物理参数的精确识别。", "limitationsZh": "无外部感知难提前应对大落差及多腿遮挡，仍发生跌倒。实机重复少且部分方法早停；图3图内与图注个别成功率不一致，本分析不把这些数字当作确定结果。训练分布需覆盖重要实际变化。", "reproductionZh": "需复现环境编码器、on-policy适应数据、奖励课程与随机化范围，严格区分潜变量更新和权重更新。真实评测应披露所有尝试及安全早停，以免夸大成功；未操作硬件。", "experimentType": "both", "robots": [ "Unitree A1" ], "evidenceNotes": [ { "section/page": "§III–IV", "note": "50步历史、8维潜变量、100/10Hz异步与A1硬件。" }, { "section/page": "Table II; §V; Figure 3", "note": "三种子千回合、实机早停，以及图内/图注不一致。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "RaiSim中改变摩擦、载荷、重心、电机能力和地形，测试范围更宽且回合内变化更频繁；三训练种子各1000回合。真实全部采用Unitree A1，比较原控制器、无适应和RMA，室内多数条件五次，严重失败时提前停在两次。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2106.14405", "title": "Habitat 2.0: Training Home Assistants to Rearrange their Habitat", "date": "2021-06-28", "url": "https://arxiv.org/abs/2106.14405v2", "titleZh": "Habitat 2.0：训练能重排居家环境的家庭助手", "abstractZh": "引入可交互ReplicaCAD场景与程序化杂物布局；以局部物理/渲染和交错执行提高速度，构建从抓取到长时序居家重排的基准。\nFetch配头部及手臂RGB-D相机，测试未见杂物布局、对象和容器，并比较端到端RL与任务规划加低层技能的组合。", "summary": "引入可交互ReplicaCAD场景与程序化杂物布局；以局部物理/渲染和交错执行提高速度，构建从抓取到长时序居家重排的基准。", "experimentType": "sim", "robots": [ "Fetch移动操作机器人（仿真）" ], "tags": [ "数据集与基准", "导航与建图", "操作与抓取" ], "limitations": "ReplicaCAD布局来自美国公寓，文化和场景覆盖有限；仿真训练环仍受环境同步与资源重载影响。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "foundation", "tierNote": "编辑分类：移动操作与居家重排的仿真和场景基础设施，适合理解完整系统评估。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2106.14405v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2106.14405v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/facebookresearch/habitat-lab", "note": "已核验官方代码与文档；未运行复现；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/facebookresearch/habitat-lab", "category": "数据集与基准 / 导航与建图", "year": 2021, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "README已核实训练/评估、Pick环境、人与仿真交互及Habitat2/3论文引用；当前主干与原论文应锁定不同版本。 本次未运行训练、复现实验或实机控制。", "directions": [ "数据集与基准", "导航与建图", "操作与抓取" ], "robotFilters": [ "Fetch移动操作机器人（仿真）" ], "original": { "id": "arxiv-2106.14405", "originalSourceUrl": "https://arxiv.org/pdf/2106.14405v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2106.14405v2", "finalDownloadUrl": "https://arxiv.org/pdf/2106.14405v2", "sha256": "92f9c30479fe4203f656c218fa228c7e5fd71f23a9821b3afa47e06dbe5ab317", "bytes": 45408472, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:09.985654+00:00", "publicationDate": "2021/06/28", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2106.14405v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2106.14405v2", "pages": 32, "sourceIdentityStatus": "title_match", "sourceTitle": "Habitat 2.0: Training Home Assistants to Rearrange their Habitat", "sourceVersion": "2106.14405v2", "sourceVersionDate": "2022/07/01", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2106.14405v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2106.14405" }, "originalAnalysis": { "id": "arxiv-2106.14405", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2106.14405v2", "sourceVersion": "2106.14405v2", "sectionsRead": [ "PDF第7页选段（提取文本L584–601）", "PDF第9页选段（提取文本L713–726）", "PDF第12页选段（提取文本L1087–1104）", "PDF第4–5页选段（提取文本L288–328）", "PDF第11–12页选段（提取文本L899–935）", "PDF第12页选段（提取文本L980–1035）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "引入可交互ReplicaCAD场景与程序化杂物布局；以局部物理/渲染和交错执行提高速度，构建从抓取到长时序居家重排的基准。", "experimentsZh": "Fetch配头部及手臂RGB-D相机，测试未见杂物布局、对象和容器，并比较端到端RL与任务规划加低层技能的组合。", "resultsZh": "论文强调简单抓取与完整家庭重排的能力差距；跨技能执行、感知和规划需要联合评估，而非只看单一抓取分数。", "limitationsZh": "ReplicaCAD布局来自美国公寓，文化和场景覆盖有限；仿真训练环仍受环境同步与资源重载影响。", "reproductionZh": "区分真实机器人型号的仿真资产与实机实验；明确是否使用理想导航、抓取简化及真值状态，锁定ReplicaCAD资源条款。", "experimentType": "sim", "robots": [ "Fetch移动操作机器人（仿真）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "引入可交互ReplicaCAD场景与程序化杂物布局；以局部物理/渲染和交错执行提高速度，构建从抓取到长时序居家重排的基准。" }, { "section": "原文实验与结果相关选段", "note": "Fetch配头部及手臂RGB-D相机，测试未见杂物布局、对象和容器，并比较端到端RL与任务规划加低层技能的组合。 论文强调简单抓取与完整家庭重排的能力差距；跨技能执行、感知和规划需要联合评估，而非只看单一抓取分数。" }, { "section": "局限与复现条件", "note": "ReplicaCAD布局来自美国公寓，文化和场景覆盖有限；仿真训练环仍受环境同步与资源重载影响。" } ], "originalSha256": "92f9c30479fe4203f656c218fa228c7e5fd71f23a9821b3afa47e06dbe5ab317", "readingScopeSha256": "deefb211099a52c76657487379fb93e80dc4dd66af236f4216bd0849f9b02987", "textExtractionSha256": "5668c8cc49688e14e4548adca9b21e199b7480b649e4d55e18741635b448c6aa" }, "experimentNote": "Fetch配头部及手臂RGB-D相机，测试未见杂物布局、对象和容器，并比较端到端RL与任务规划加低层技能的组合。", "contribution": "引入可交互ReplicaCAD场景与程序化杂物布局；以局部物理/渲染和交错执行提高速度，构建从抓取到长时序居家重排的基准。", "whyUseful": "区分真实机器人型号的仿真资产与实机实验；明确是否使用理想导航、抓取简化及真值状态，锁定ReplicaCAD资源条款。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "核心MIT；Habitat-Sim、ReplicaCAD等数据资产另有条款。", "codeReleaseScope": "已核验官方代码与文档；未运行复现" }, { "id": "arxiv-2104.02180", "title": "AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control", "titleZh": "AMP：面向风格化物理角色控制的对抗运动先验", "shortTitle": "AMP", "date": "2021-04-05", "datePrecision": "day", "year": 2021, "url": "https://arxiv.org/abs/2104.02180", "paperUrl": "https://arxiv.org/abs/2104.02180", "codeUrl": "https://github.com/xbpeng/DeepMimic", "summary": "AMP用相邻状态的对抗判别器产生动作风格奖励，与目标方向、速度或操作任务奖励相加，再用PPO优化控制。最小二乘GAN和梯度惩罚稳定判别器，输入包含局部姿态及速度，不需要参考动作相位或手工选择动作片段。", "abstractZh": "AMP用相邻状态的对抗判别器产生动作风格奖励，与目标方向、速度或操作任务奖励相加，再用PPO优化控制。最小二乘GAN和梯度惩罚稳定判别器，输入包含局部姿态及速度，不需要参考动作相位或手工选择动作片段。\nBullet仿真34自由度人形、59自由度霸王龙及64自由度狗，30Hz策略、1.2kHz物理步；测试行走转向、起身、踢击、盘球、跨障碍和跳石。三模型种子各32回合，另比较单动作模仿和无速度、无梯度惩罚等消融。", "category": "运动控制 / 对抗运动先验", "tags": [ "对抗模仿", "运动先验", "风格奖励", "强化学习" ], "directions": [ "运动控制", "基础理论", "强化学习", "模仿学习" ], "tier": "foundation", "experimentType": "sim", "robots": [], "robotNote": "实验为虚拟物理角色；不把后续真实四足或人形AMP论文的硬件归到本篇。", "codeStatus": "open", "status": "作者实现公开；历史仓库已弃用并推荐MimicKit。", "trainingNote": "交替训练运动判别器和强化学习策略，叠加风格回报与任务回报；原仓库包含AMP训练入口。", "contribution": "AMP用相邻状态的对抗判别器产生动作风格奖励，与目标方向、速度或操作任务奖励相加，再用PPO优化控制。最小二乘GAN和梯度惩罚稳定判别器，输入包含局部姿态及速度，不需要参考动作相位或手工选择动作片段。", "whyUseful": "先用短片段检查局部坐标、速度特征及终止条件，再加入多技能库并测模式覆盖。分别报告任务回报、动作相似度和跌倒，固定1–3亿样本预算；不要把后续机器人AMP应用当作原论文硬件证据。", "limitations": "全部是动画物理角色，未验证真实人形机器人；判别器可能模式坍塌而忽略动作库的大部分技能。每个策略的运动先验仍从头训练，跨任务复用未解决；任务训练分布也会影响看似任务无关的先验。", "evidence": [ { "url": "https://arxiv.org/abs/2104.02180", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://xbpeng.github.io/projects/AMP/index.html", "note": "作者页面明确判别式风格回报、无需片段顺序标注与仿真范围。" }, { "url": "https://github.com/xbpeng/DeepMimic", "note": "官方README列出AMP及训练依赖，提示旧仓库弃用。" }, { "url": "https://arxiv.org/pdf/2104.02180", "note": "5–6; 8.1：状态转移判别器、LSGAN/PPO及模拟角色与频率。" }, { "url": "https://arxiv.org/pdf/2104.02180", "note": "Tables 1–3; 9：任务与姿态误差区别，模式坍塌和先验复用限制。" }, { "url": "https://arxiv.org/pdf/2104.02180v2", "note": "5–6; 8.1：状态转移判别器、LSGAN/PPO及模拟角色与频率。" }, { "url": "https://arxiv.org/pdf/2104.02180v2", "note": "Tables 1–3; 9：任务与姿态误差区别，模式坍塌和先验复用限制。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2021：运动控制从显式轨迹跟踪扩展到数据驱动的分布式运动先验。", "projectUrl": null, "robotFilters": [], "original": { "id": "arxiv-2104.02180", "originalSourceUrl": "https://arxiv.org/pdf/2104.02180v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2104.02180v2", "finalDownloadUrl": "https://arxiv.org/pdf/2104.02180", "sha256": "93d9e8e40b3908d65e8b88ae3dcbe8d4cd69071d13f3234ad927462477472a29", "bytes": 8715012, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:54.872931+00:00", "publicationDate": "2021/04/05", "publicationDateSourceUrl": "https://arxiv.org/abs/2104.02180", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/2104.02180", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2104.02180", "pages": 20, "sourceIdentityStatus": "title_match", "sourceTitle": "AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control", "sourceVersion": "2104.02180v2", "sourceVersionDate": "2022/05/12", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.362357+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2104.02180" }, "originalAnalysis": { "id": "arxiv-2104.02180", "sourceUrl": "https://arxiv.org/pdf/2104.02180v2", "sectionsRead": [ "5.1–5.4", "6–8", "9", "Tables 1–3" ], "methodsZh": "AMP用相邻状态的对抗判别器产生动作风格奖励，与目标方向、速度或操作任务奖励相加，再用PPO优化控制。最小二乘GAN和梯度惩罚稳定判别器，输入包含局部姿态及速度，不需要参考动作相位或手工选择动作片段。", "experimentsZh": "Bullet仿真34自由度人形、59自由度霸王龙及64自由度狗，30Hz策略、1.2kHz物理步；测试行走转向、起身、踢击、盘球、跨障碍和跳石。三模型种子各32回合，另比较单动作模仿和无速度、无梯度惩罚等消融。", "resultsZh": "多片段数据可自发出现按目标速度从走到跑的切换；人形方向跟踪归一化任务回报0.90。单片段表明视觉合理不等于总是更准：后空翻姿态误差0.150米，对显式跟踪0.076米，而侧翻0.124对0.191更好。", "limitationsZh": "全部是动画物理角色，未验证真实人形机器人；判别器可能模式坍塌而忽略动作库的大部分技能。每个策略的运动先验仍从头训练，跨任务复用未解决；任务训练分布也会影响看似任务无关的先验。", "reproductionZh": "先用短片段检查局部坐标、速度特征及终止条件，再加入多技能库并测模式覆盖。分别报告任务回报、动作相似度和跌倒，固定1–3亿样本预算；不要把后续机器人AMP应用当作原论文硬件证据。", "experimentType": "sim", "robots": [], "corrections": [ "原文对象为模拟人形/动物角色，不能填Unitree等后续应用硬件。" ], "evidenceNotes": [ { "note": "状态转移判别器、LSGAN/PPO及模拟角色与频率。", "section": "5–6; 8.1" }, { "note": "任务与姿态误差区别，模式坍塌和先验复用限制。", "section": "Tables 1–3; 9" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2104.02180v2", "originalSha256": "93d9e8e40b3908d65e8b88ae3dcbe8d4cd69071d13f3234ad927462477472a29" }, "experimentNote": "Bullet仿真34自由度人形、59自由度霸王龙及64自由度狗，30Hz策略、1.2kHz物理步；测试行走转向、起身、踢击、盘球、跨障碍和跳石。三模型种子各32回合，另比较单动作模仿和无速度、无梯度惩罚等消融。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2103.12352", "title": "iMAP: Implicit Mapping and Positioning in Real-Time", "date": "2021-03-23", "url": "https://arxiv.org/abs/2103.12352v2", "titleZh": "iMAP：实时隐式建图与定位", "abstractZh": "用单个在线训练的MLP将三维坐标映射到颜色和体密度，跟踪与建图并行运行；关键帧重放和误差引导像素采样缓解遗忘并减少优化成本。\nReplica合成RGB-D序列用于重建，手持Azure Kinect与TUM RGB-D用于真实图像跟踪及重建；未报告机器人导航闭环。", "summary": "用单个在线训练的MLP将三维坐标映射到颜色和体密度，跟踪与建图并行运行；关键帧重放和误差引导像素采样缓解遗忘并减少优化成本。", "experimentType": "data", "robots": [], "tags": [ "导航与建图" ], "limitations": "评测主要是房间尺度；网络不建模视角相关镜面效果，补全属于模型推断，不能当作已观测的可通行空间。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "classic", "tierNote": "编辑分类：具有代表性的在线隐式建图早期方法，用于理解神经场与传统地图的取舍。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现待核实", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2103.12352v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2103.12352v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" } ], "codeStatus": "unknown", "codeUrl": null, "category": "导航与建图", "year": 2021, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "已核验原文；本轮未独立核实官方实现、训练入口及发布范围，不能据论文声明认定代码已开放。 本次未运行训练、复现实验或实机控制。", "directions": [ "导航与建图" ], "robotFilters": [], "original": { "id": "arxiv-2103.12352", "originalSourceUrl": "https://arxiv.org/pdf/2103.12352v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2103.12352v2", "finalDownloadUrl": "https://arxiv.org/pdf/2103.12352v2", "sha256": "bd10370f5fc5eedc1ea3ebc4bd6fd7164164b3b7aad38668035b69b4038f378c", "bytes": 8480607, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:00:38.235804+00:00", "publicationDate": "2021/03/23", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2103.12352v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2103.12352v2", "pages": 10, "sourceIdentityStatus": "title_match", "sourceTitle": "iMAP: Implicit Mapping and Positioning in Real-Time", "sourceVersion": "2103.12352v2", "sourceVersionDate": "2021/09/13", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2103.12352v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2103.12352" }, "originalAnalysis": { "id": "arxiv-2103.12352", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2103.12352v2", "sourceVersion": "2103.12352v2", "sectionsRead": [ "PDF第4–5页选段（提取文本L429–446）", "PDF第5–6页选段（提取文本L487–500）", "PDF第2–3页选段（提取文本L145–170）", "PDF第3页选段（提取文本L225–244）", "PDF第8–9页选段（提取文本L919–938）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "用单个在线训练的MLP将三维坐标映射到颜色和体密度，跟踪与建图并行运行；关键帧重放和误差引导像素采样缓解遗忘并减少优化成本。", "experimentsZh": "Replica合成RGB-D序列用于重建，手持Azure Kinect与TUM RGB-D用于真实图像跟踪及重建；未报告机器人导航闭环。", "resultsZh": "证明紧凑神经网络可作为房间级在线稠密地图；论文展示比TSDF更紧凑的表示和对部分缺失观测的补全能力。", "limitationsZh": "评测主要是房间尺度；网络不建模视角相关镜面效果，补全属于模型推断，不能当作已观测的可通行空间。", "reproductionZh": "锁定关键帧窗口、像素采样与网络容量，分别核对跟踪误差和几何误差；机器人安全地图应保留观测置信度。", "experimentType": "data", "robots": [], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "用单个在线训练的MLP将三维坐标映射到颜色和体密度，跟踪与建图并行运行；关键帧重放和误差引导像素采样缓解遗忘并减少优化成本。" }, { "section": "原文实验与结果相关选段", "note": "Replica合成RGB-D序列用于重建，手持Azure Kinect与TUM RGB-D用于真实图像跟踪及重建；未报告机器人导航闭环。 证明紧凑神经网络可作为房间级在线稠密地图；论文展示比TSDF更紧凑的表示和对部分缺失观测的补全能力。" }, { "section": "局限与复现条件", "note": "评测主要是房间尺度；网络不建模视角相关镜面效果，补全属于模型推断，不能当作已观测的可通行空间。" } ], "originalSha256": "bd10370f5fc5eedc1ea3ebc4bd6fd7164164b3b7aad38668035b69b4038f378c", "readingScopeSha256": "fcc91f896bf192f5b8f9196c809ad8f6b3496fb58883d7369dd8e763973d8b61", "textExtractionSha256": "145c614bc9d59aa483b67e7a475f4260370b05f978cb5ce09f6784bdabb40590" }, "experimentNote": "Replica合成RGB-D序列用于重建，手持Azure Kinect与TUM RGB-D用于真实图像跟踪及重建；未报告机器人导航闭环。", "contribution": "用单个在线训练的MLP将三维坐标映射到颜色和体密度，跟踪与建图并行运行；关键帧重放和误差引导像素采样缓解遗忘并减少优化成本。", "whyUseful": "锁定关键帧窗口、像素采样与网络容量，分别核对跟踪误差和几何误差；机器人安全地图应保留观测置信度。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": null, "codeReleaseScope": "未独立核实" }, { "id": "arxiv-2010.14406", "title": "Transporter Networks: Rearranging the Visual World for Robotic Manipulation", "titleZh": "Transporter网络：为机器人操作重排视觉世界", "shortTitle": "Transporter Networks", "date": "2020-10-27", "datePrecision": "day", "year": 2020, "url": "https://arxiv.org/abs/2010.14406", "codeUrl": "https://github.com/google-research/ravens", "category": "模仿学习", "tags": [ "空间等变性", "抓放", "Ravens" ], "tier": "classic", "summary": "Transporter先用全卷积网络选抓取点，再裁切该点附近的视觉特征，经平移旋转后与整幅特征互相关，预测条件放置位姿。RGB-D转俯视颜色/高度栅格，把动作与像素坐标对齐；交叉熵拟合示范，利用空间等变性表达多…", "abstractZh": "Transporter先用全卷积网络选抓取点，再裁切该点附近的视觉特征，经平移旋转后与整幅特征互相关，预测条件放置位姿。RGB-D转俯视颜色/高度栅格，把动作与像素坐标对齐；交叉熵拟合示范，利用空间等变性表达多种合法抓放组合，也可用两位姿原语推扫物体。\nRavens含十项PyBullet任务，每任务用1、10、100或1000条脚本示范单独训练，在100个未见测试配置评价最佳验证模型；对照Form2Fit、图像MLP及真状态MLP。真实UR5系列用13名操作者采集8141次抓放和6759次推扫，分别做瓶装配与围棋子清扫。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "Universal Robots UR5/UR5e" ], "codeStatus": "open", "status": "官方Ravens训练、评测和环境代码公开；仓库已归档。", "trainingNote": "示范监督学习；仿真可通过脚本专家生成数据。", "whyUseful": "复现需160×320俯视网格、36个平面角度、统一SE(2)增强、原语及成功度计算。正文和图标UR5e而硬件段也称UR5，应保留系列层级并核查具体配置；训练示范条数和真实动作转移数不可直接对比。", "contribution": "Transporter先用全卷积网络选抓取点，再裁切该点附近的视觉特征，经平移旋转后与整幅特征互相关，预测条件放置位姿。RGB-D转俯视颜色/高度栅格，把动作与像素坐标对齐；交叉熵拟合示范，利用空间等变性表达多种合法抓放组合，也可用两位姿原语推扫物体。", "limitations": "每个任务单独模型，依赖相机与机器人标定及可执行运动原语；高频力矩/力控制未解决。本文真实模型用实机示范训练，仿真用于受控比较，并非直接sim-to-real。无记忆策略在部分可观测场景可能受限。", "caveats": "官方仓库公开不代表自带完整硬件部署；UR5型号来自真实实验附录。", "evidence": [ { "url": "https://arxiv.org/abs/2010.14406", "note": "首发日期、方法与仿真/真实实验。" }, { "url": "https://arxiv.org/pdf/2010.14406", "note": "附录真实实验使用两个UR5工作站。" }, { "url": "https://transporternets.github.io/", "note": "项目页列出代码、模型与数据。" }, { "url": "https://github.com/google-research/ravens", "note": "官方仓库及Apache-2.0许可；显示已归档。" }, { "url": "https://arxiv.org/pdf/2010.14406v3", "note": "3.1–3.3：Pick-conditioned cross-correlation;one task/model;36 orientations;pixel cross-entropy." }, { "url": "https://arxiv.org/pdf/2010.14406v3", "note": "4.1 表2/4：100 unseen test runs;partial credit for multistep tasks;6DoF score91 with1000 demos." }, { "url": "https://arxiv.org/pdf/2010.14406v3", "note": "4.2/附录6.14：Real data13 operators,8141 pick/place transitions,6759 pushes;manual robot-action labels;98.9/98.3 reported." }, { "url": "https://arxiv.org/pdf/2010.14406v3", "note": "4.1/5：Explicitly not sim-to-real;camera calibration sensitivity;force/torque integration unresolved." } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "操作与抓取", "模仿学习" ], "robotFilters": [ "Universal Robots UR5" ], "verificationNote": "2026-10-04核对论文、项目页及官方仓库。", "original": { "id": "arxiv-2010.14406", "originalSourceUrl": "https://arxiv.org/pdf/2010.14406v3", "versionedOriginalUrl": "https://arxiv.org/pdf/2010.14406v3", "finalDownloadUrl": "https://arxiv.org/pdf/2010.14406", "sha256": "3a5e78b5c8881863845c46bda0bf8c9071aa79a59a6f9a9f4d8732690482b228", "bytes": 3781112, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:54.796116+00:00", "publicationDate": "2020/10/27", "publicationDateSourceUrl": "https://arxiv.org/abs/2010.14406", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2010.14406", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2010.14406", "pages": 23, "sourceIdentityStatus": "title_match", "sourceTitle": "Transporter Networks: Rearranging the Visual World for Robotic Manipulation", "sourceVersion": "2010.14406v3", "sourceVersionDate": "2022/01/05", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.366398+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2010.14406" }, "originalAnalysis": { "id": "arxiv-2010.14406", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739406+00:00", "sourceUrl": "https://arxiv.org/pdf/2010.14406v3", "sectionsRead": [ "3.1–3.3 抓取条件搬运特征、网络与示范学习", "4.1/表2–4 Ravens和SE(3)实验", "4.2/5 实机和局限", "附录6.5/6.14 指标与硬件" ], "methodsZh": "Transporter先用全卷积网络选抓取点，再裁切该点附近的视觉特征，经平移旋转后与整幅特征互相关，预测条件放置位姿。RGB-D转俯视颜色/高度栅格，把动作与像素坐标对齐；交叉熵拟合示范，利用空间等变性表达多种合法抓放组合，也可用两位姿原语推扫物体。", "experimentsZh": "Ravens含十项PyBullet任务，每任务用1、10、100或1000条脚本示范单独训练，在100个未见测试配置评价最佳验证模型；对照Form2Fit、图像MLP及真状态MLP。真实UR5系列用13名操作者采集8141次抓放和6759次推扫，分别做瓶装配与围棋子清扫。", "resultsZh": "作者报告平面插块一条示范已达100分，但积木金字塔1000条仍为78.2分，6DoF插块1000条为91分。真实装配/清扫测试表现为98.9%/98.3%；须注意任务指标允许部分完成得分，不能把所有百分数一概当整段成功率，且实机测试分母未充分给出。", "limitationsZh": "每个任务单独模型，依赖相机与机器人标定及可执行运动原语；高频力矩/力控制未解决。本文真实模型用实机示范训练，仿真用于受控比较，并非直接sim-to-real。无记忆策略在部分可观测场景可能受限。", "reproductionZh": "复现需160×320俯视网格、36个平面角度、统一SE(2)增强、原语及成功度计算。正文和图标UR5e而硬件段也称UR5，应保留系列层级并核查具体配置；训练示范条数和真实动作转移数不可直接对比。", "evidenceNotes": [ { "section/page": "3.1–3.3", "note": "Pick-conditioned cross-correlation;one task/model;36 orientations;pixel cross-entropy." }, { "section/page": "4.1 表2/4", "note": "100 unseen test runs;partial credit for multistep tasks;6DoF score91 with1000 demos." }, { "section/page": "4.2/附录6.14", "note": "Real data13 operators,8141 pick/place transitions,6759 pushes;manual robot-action labels;98.9/98.3 reported." }, { "section/page": "4.1/5", "note": "Explicitly not sim-to-real;camera calibration sensitivity;force/torque integration unresolved." } ], "experimentType": "both", "robots": [ "Universal Robots UR5/UR5e" ], "corrections": [ "仿真模型并非直接迁移到实机；真实训练使用人类遥操作数据。", "原文UR5/UR5e命名混用，避免武断细分。" ], "sourceVersion": "2010.14406v3", "originalSha256": "3a5e78b5c8881863845c46bda0bf8c9071aa79a59a6f9a9f4d8732690482b228" }, "experimentNote": "Ravens含十项PyBullet任务，每任务用1、10、100或1000条脚本示范单独训练，在100个未见测试配置评价最佳验证模型；对照Form2Fit、图像MLP及真状态MLP。真实UR5系列用13名操作者采集8141次抓放和6759次推扫，分别做瓶装配与围棋子清扫。", "analysisVerifiedAt": "2026-10-04T13:47:58.739406+00:00" }, { "id": "arxiv-2010.02193", "title": "Mastering Atari with Discrete World Models", "titleZh": "利用离散世界模型掌握 Atari", "shortTitle": "DreamerV2", "date": "2020-10-05", "year": 2020, "url": "https://arxiv.org/abs/2010.02193", "paperUrl": "https://arxiv.org/abs/2010.02193", "projectUrl": "https://danijar.com/project/dreamerv2/", "codeUrl": "https://github.com/danijar/dreamerv2", "summary": "DreamerV2以离散分类潜变量取代高斯状态，用直通梯度训练RSSM；KL balancing分别调节先验拟合和后验正则。世界模型重建图像并预测奖励/终止，actor-critic仅在潜在想象中优化，离散动作…", "abstractZh": "DreamerV2以离散分类潜变量取代高斯状态，用直通梯度训练RSSM；KL balancing分别调节先验拟合和后验正则。世界模型重建图像并预测奖励/终止，actor-critic仅在潜在想象中优化，离散动作结合REINFORCE等梯度估计。\n55个Atari游戏各自训练独立智能体，200M环境步、动作重复4、sticky actions、完整动作集，不使用生命信息或帧堆叠。单GPU单环境，与Dopamine版本IQN、Rainbow等比较；消融潜变量、KL、图像/奖励梯度。", "category": "世界模型 / 想象强化学习", "tags": [ "DreamerV2", "世界模型", "强化学习" ], "directions": [ "世界模型", "强化学习" ], "tier": "classic", "experimentType": "sim", "experimentNote": "55个Atari游戏各自训练独立智能体，200M环境步、动作重复4、sticky actions、完整动作集，不使用生命信息或帧堆叠。单GPU单环境，与Dopamine版本IQN、Rainbow等比较；消融潜变量、KL、图像/奖励梯度。", "robots": [], "robotFilters": [], "robotNote": "无实机；仿真形态不等同于已验证的商业机器人型号。", "codeStatus": "open", "status": "训练代码已开源", "trainingStatus": "训练代码已开源", "trainingNote": "官方 TensorFlow 2 仓库提供训练代码和 55 个游戏分数，MIT 许可。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "MIT（代码）", "contribution": "DreamerV2以离散分类潜变量取代高斯状态，用直通梯度训练RSSM；KL balancing分别调节先验拟合和后验正则。世界模型重建图像并预测奖励/终止，actor-critic仅在潜在想象中优化，离散动作结合REINFORCE等梯度估计。", "whyUseful": "复现需锁定55游戏列表、sticky动作、环境帧/决策步换算和归一化参考，保留逐游戏结果。消融用略早版本，不能将表2直接当表1同配置的精确差分；公开实现仍需独立运行核验。", "limitations": "这是游戏与补充连续控制研究，没有真实机器人证据。“人类水平”是特定归一化聚合而非每款游戏都超过人类；小目标如Video Pinball单像素球仍困难，均值也受参考分数和截断规则影响。", "caveats": "实验为游戏与仿真人形，无真实人形机器人部署；难探索任务仍具挑战。", "evidence": [ { "url": "https://arxiv.org/abs/2010.02193", "note": "首发日期与摘要。" }, { "url": "https://arxiv.org/html/2010.02193v4", "note": "§2 离散潜变量、KL 平衡，附录模拟人形。" }, { "url": "https://github.com/danijar/dreamerv2", "note": "官方实现、分数及 MIT。" }, { "url": "https://arxiv.org/pdf/2010.02193", "note": "§3 Experimental setup：55个独立游戏，200M、sticky动作、单GPU。" }, { "url": "https://arxiv.org/pdf/2010.02193", "note": "Table 1：Gamer median2.15；clipped record mean0.28。" }, { "url": "https://arxiv.org/pdf/2010.02193", "note": "Fig.5/Table 2：消融使用较早版本；VideoPinball小目标失败见§3.1。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2020：离散潜状态推动世界模型达到强 Atari 表现，并支持连续控制。", "freshness": "经典工作", "original": { "id": "arxiv-2010.02193", "originalSourceUrl": "https://arxiv.org/pdf/2010.02193v4", "versionedOriginalUrl": "https://arxiv.org/pdf/2010.02193v4", "finalDownloadUrl": "https://arxiv.org/pdf/2010.02193", "sha256": "037dce26de328dd86ce4ab71dfb54f44513fef49a02f2995a1c05848ccef158f", "bytes": 5261247, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:39:58.472997+00:00", "publicationDate": "2020/10/05", "publicationDateSourceUrl": "https://arxiv.org/abs/2010.02193", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2010.02193", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2010.02193", "pages": 26, "sourceIdentityStatus": "title_match", "sourceTitle": "Mastering Atari with Discrete World Models", "sourceVersion": "2010.02193v4", "sourceVersionDate": "2022/02/12", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.371141+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2010.02193" }, "originalAnalysis": { "id": "arxiv-2010.02193", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090147+00:00", "sourceUrl": "https://arxiv.org/pdf/2010.02193", "sectionsRead": [ "§2", "§3.1–3.3", "Tables 1–2", "Fig.5", "§5" ], "methodsZh": "DreamerV2以离散分类潜变量取代高斯状态，用直通梯度训练RSSM；KL balancing分别调节先验拟合和后验正则。世界模型重建图像并预测奖励/终止，actor-critic仅在潜在想象中优化，离散动作结合REINFORCE等梯度估计。", "experimentsZh": "55个Atari游戏各自训练独立智能体，200M环境步、动作重复4、sticky actions、完整动作集，不使用生命信息或帧堆叠。单GPU单环境，与Dopamine版本IQN、Rainbow等比较；消融潜变量、KL、图像/奖励梯度。", "resultsZh": "作者表1游戏玩家归一化中位2.15，对Rainbow1.47；截断世界纪录归一化均值0.28，对IQN0.21。离散状态与KL平衡有益，移除图像梯度性能明显下滑；不同聚合方式会改变基线相对排名。", "limitationsZh": "这是游戏与补充连续控制研究，没有真实机器人证据。“人类水平”是特定归一化聚合而非每款游戏都超过人类；小目标如Video Pinball单像素球仍困难，均值也受参考分数和截断规则影响。", "reproductionZh": "复现需锁定55游戏列表、sticky动作、环境帧/决策步换算和归一化参考，保留逐游戏结果。消融用略早版本，不能将表2直接当表1同配置的精确差分；公开实现仍需独立运行核验。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "§3 Experimental setup", "note": "55个独立游戏，200M、sticky动作、单GPU。" }, { "section/page": "Table 1", "note": "Gamer median2.15；clipped record mean0.28。" }, { "section/page": "Fig.5/Table 2", "note": "消融使用较早版本；VideoPinball小目标失败见§3.1。" } ] }, "analysisVerifiedAt": "2026-10-04T13:50:35.090147+00:00" }, { "id": "arxiv-2009.12293", "title": "robosuite: A Modular Simulation Framework and Benchmark for Robot Learning", "date": "2020-09-25", "url": "https://arxiv.org/abs/2009.12293v3", "titleZh": "robosuite：模块化机器人学习仿真框架与基准", "abstractZh": "把机器人、末端、底座、物体和场景模块组装成MuJoCo模型，提供多模态传感器、遥操作和任务空间/关节空间控制器。\n所读v3文档介绍robosuite 1.5；基准以Panda和Sawyer、不同控制器训练SAC，使用低维物理状态并报告五个随机种子。", "summary": "把机器人、末端、底座、物体和场景模块组装成MuJoCo模型，提供多模态传感器、遥操作和任务空间/关节空间控制器。", "experimentType": "sim", "robots": [ "Franka Panda（仿真）", "Rethink Sawyer（仿真）" ], "tags": [ "数据集与基准", "操作与抓取", "强化学习" ], "limitations": "基准结果限于所选任务、状态观测和控制器；支持某机器人模型不代表其实机接口或动力学已被验证。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "foundation", "tierNote": "编辑分类：模块化操作仿真与控制器基础设施，常用于搭建强化学习和模仿学习实验。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/2009.12293v3", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/2009.12293v3", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/ARISE-Initiative/robosuite", "note": "已核验官方代码与文档；未运行复现；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/ARISE-Initiative/robosuite", "category": "数据集与基准 / 操作与抓取", "year": 2020, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "README和许可证已核实仿真环境、控制器、示教采集接口；策略训练常由robomimic等配套库完成。 本次未运行训练、复现实验或实机控制。", "directions": [ "数据集与基准", "操作与抓取", "强化学习" ], "robotFilters": [ "Franka Panda（仿真）", "Rethink Sawyer（仿真）" ], "original": { "id": "arxiv-2009.12293", "originalSourceUrl": "https://arxiv.org/pdf/2009.12293v3", "versionedOriginalUrl": "https://arxiv.org/pdf/2009.12293v3", "finalDownloadUrl": "https://arxiv.org/pdf/2009.12293v3", "sha256": "3da1cc7ba84ee487a90d7013325dd6045eea7c2920bc50cdb8ff37db67426fd4", "bytes": 18253350, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T15:01:24.788140+00:00", "publicationDate": "2020/09/25", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2009.12293v3", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2009.12293v3", "pages": 19, "sourceIdentityStatus": "title_match", "sourceTitle": "robosuite: A Modular Simulation Framework and Benchmark for Robot Learning", "sourceVersion": "2009.12293v3", "sourceVersionDate": "2025/01/18", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/2009.12293v3", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2009.12293" }, "originalAnalysis": { "id": "arxiv-2009.12293", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/2009.12293v3", "sourceVersion": "2009.12293v3", "sectionsRead": [ "PDF第11–12页选段（提取文本L618–635）", "PDF第15页选段（提取文本L756–769）", "PDF第4页选段（提取文本L148–177）", "PDF第15–16页选段（提取文本L769–785）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "把机器人、末端、底座、物体和场景模块组装成MuJoCo模型，提供多模态传感器、遥操作和任务空间/关节空间控制器。", "experimentsZh": "所读v3文档介绍robosuite 1.5；基准以Panda和Sawyer、不同控制器训练SAC，使用低维物理状态并报告五个随机种子。", "resultsZh": "在抬块与开门对照中，仅改变控制器就明显影响学习效率，任务空间控制收敛更快；框架配置本身是重要实验变量。", "limitationsZh": "基准结果限于所选任务、状态观测和控制器；支持某机器人模型不代表其实机接口或动力学已被验证。", "reproductionZh": "必须注明论文v3与robosuite版本，保留控制器参数、MuJoCo版本、奖励尺度及观测配置，避免将最新功能归到2020首版。", "experimentType": "sim", "robots": [ "Franka Panda（仿真）", "Rethink Sawyer（仿真）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "把机器人、末端、底座、物体和场景模块组装成MuJoCo模型，提供多模态传感器、遥操作和任务空间/关节空间控制器。" }, { "section": "原文实验与结果相关选段", "note": "所读v3文档介绍robosuite 1.5；基准以Panda和Sawyer、不同控制器训练SAC，使用低维物理状态并报告五个随机种子。 在抬块与开门对照中，仅改变控制器就明显影响学习效率，任务空间控制收敛更快；框架配置本身是重要实验变量。" }, { "section": "局限与复现条件", "note": "基准结果限于所选任务、状态观测和控制器；支持某机器人模型不代表其实机接口或动力学已被验证。" } ], "originalSha256": "3da1cc7ba84ee487a90d7013325dd6045eea7c2920bc50cdb8ff37db67426fd4", "readingScopeSha256": "4f1530a61a8bb97b36c4fea09a3558cff8652e2227c464b4896e2b3956071914", "textExtractionSha256": "0f7dd7990d8ffd76821bed0defb34c4b0620ed04e9b5e170c217e621f8a5ed27" }, "experimentNote": "所读v3文档介绍robosuite 1.5；基准以Panda和Sawyer、不同控制器训练SAC，使用低维物理状态并报告五个随机种子。", "contribution": "把机器人、末端、底座、物体和场景模块组装成MuJoCo模型，提供多模态传感器、遥操作和任务空间/关节空间控制器。", "whyUseful": "必须注明论文v3与robosuite版本，保留控制器参数、MuJoCo版本、奖励尺度及观测配置，避免将最新功能归到2020首版。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "核心MIT；LICENSE另注明MuJoCo相关代码为Apache-2.0。", "codeReleaseScope": "已核验官方代码与文档；未运行复现" }, { "id": "arxiv-2007.11898", "title": "ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM", "titleZh": "ORB-SLAM3：精确的视觉、视觉惯性与多地图开源 SLAM 库", "date": "2020-07-23", "datePrecision": "day", "year": 2020, "url": "https://arxiv.org/abs/2007.11898", "paperUrl": "https://arxiv.org/abs/2007.11898", "codeUrl": "https://github.com/UZ-SLAMLab/ORB_SLAM3", "summary": "在特征点关键帧SLAM中加入MAP视觉惯性初始化及Atlas多地图，追踪失败可新建地图，再通过高召回地点识别与局部束调整融合。支持单目、双目、RGB-D、针孔/鱼眼和惯性模式，复用中期地图关联改善仅短窗口里程计…", "abstractZh": "在特征点关键帧SLAM中加入MAP视觉惯性初始化及Atlas多地图，追踪失败可新建地图，再通过高召回地点识别与局部束调整融合。支持单目、双目、RGB-D、针孔/鱼眼和惯性模式，复用中期地图关联改善仅短窗口里程计的漂移。\n在EuRoC与TUM-VI真实传感器序列比较多配置单会话和多会话定位；报告全轨迹RMS ATE、尺度误差和处理时间，按协议取多次运行中位数。多会话依次处理同地多序列并一次全局对齐，不是每段单独消除漂移。", "category": "视觉惯性 SLAM / 多地图", "tags": [ "ORB-SLAM3", "视觉惯性", "多地图", "重定位", "真实数据评测" ], "directions": [ "导航与建图" ], "tier": "classic", "experimentType": "data", "robots": [], "robotFilters": [], "codeStatus": "open", "status": "实现已开源", "trainingNote": "无需神经策略训练；官方仓库提供相机/IMU 标定指导和 EuRoC、TUM-VI 示例。代码按 GPLv3 发布。", "contribution": "在特征点关键帧SLAM中加入MAP视觉惯性初始化及Atlas多地图，追踪失败可新建地图，再通过高召回地点识别与局部束调整融合。支持单目、双目、RGB-D、针孔/鱼眼和惯性模式，复用中期地图关联改善仅短窗口里程计的漂移。", "whyUseful": "需严格核对相机与IMU标定、图像均衡、特征数量、天空远点过滤、对齐自由度及失败序列纳入方式。没有神经策略训练；本次只读论文，未运行公开库。", "limitations": "单目纯视觉允许7自由度尺度对齐，双目/惯性6自由度，不能直接比较裸数判胜负。表内部分外部结果使用不同真值或关键帧轨迹；远景、低纹理和弱惯性激励仍会失败或漂移。", "license": "GPL-3.0", "evidence": [ { "url": "https://arxiv.org/abs/2007.11898", "note": "首版日期、视觉惯性 MAP 初始化、多地图与真实数据集实验。" }, { "url": "https://github.com/UZ-SLAMLab/ORB_SLAM3", "note": "官方 GPLv3 仓库，提供 EuRoC/TUM-VI 配置及标定材料。" }, { "url": "https://arxiv.org/pdf/2007.11898", "note": "§III–V：Atlas、多地图融合与MAP初始化。" }, { "url": "https://arxiv.org/pdf/2007.11898", "note": "Tables II–V; §VII-D：ATE数字、对齐/真值脚注与30–40FPS；真实记录数据评测。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "把紧耦合惯性估计和多地图恢复整合进成熟特征型 SLAM 系统。", "experimentNote": "在EuRoC与TUM-VI真实传感器序列比较多配置单会话和多会话定位；报告全轨迹RMS ATE、尺度误差和处理时间，按协议取多次运行中位数。多会话依次处理同地多序列并一次全局对齐，不是每段单独消除漂移。", "projectUrl": null, "shortTitle": "", "original": { "id": "arxiv-2007.11898", "originalSourceUrl": "https://arxiv.org/pdf/2007.11898v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2007.11898v2", "finalDownloadUrl": "https://arxiv.org/pdf/2007.11898", "sha256": "4d119517be7565c9651fbf1e36a330d605b7a534d738558858c96feca7b99156", "bytes": 5049669, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:40:02.753784+00:00", "publicationDate": "2020/07/23", "publicationDateSourceUrl": "https://arxiv.org/abs/2007.11898", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2007.11898", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2007.11898", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM", "sourceVersion": "2007.11898v2", "sourceVersionDate": "2021/04/23", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.375884+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2007.11898" }, "originalAnalysis": { "id": "arxiv-2007.11898", "sourceUrl": "https://arxiv.org/pdf/2007.11898", "sectionsRead": [ "§III–VI", "§VII Tables II–VII", "§VIII" ], "methodsZh": "在特征点关键帧SLAM中加入MAP视觉惯性初始化及Atlas多地图，追踪失败可新建地图，再通过高召回地点识别与局部束调整融合。支持单目、双目、RGB-D、针孔/鱼眼和惯性模式，复用中期地图关联改善仅短窗口里程计的漂移。", "experimentsZh": "在EuRoC与TUM-VI真实传感器序列比较多配置单会话和多会话定位；报告全轨迹RMS ATE、尺度误差和处理时间，按协议取多次运行中位数。多会话依次处理同地多序列并一次全局对齐，不是每段单独消除漂移。", "resultsZh": "作者报告EuRoC单目惯性平均ATE0.043米、双目惯性0.035米；TUM-VI小房间双目惯性平均0.009米。部分长室外序列仍数十米误差。主线程约30–40帧/秒；这些定位基准成绩不等于机器人导航任务成功。", "limitationsZh": "单目纯视觉允许7自由度尺度对齐，双目/惯性6自由度，不能直接比较裸数判胜负。表内部分外部结果使用不同真值或关键帧轨迹；远景、低纹理和弱惯性激励仍会失败或漂移。", "reproductionZh": "需严格核对相机与IMU标定、图像均衡、特征数量、天空远点过滤、对齐自由度及失败序列纳入方式。没有神经策略训练；本次只读论文，未运行公开库。", "experimentType": "data", "robots": [], "evidenceNotes": [ { "section/page": "§III–V", "note": "Atlas、多地图融合与MAP初始化。" }, { "section/page": "Tables II–V; §VII-D", "note": "ATE数字、对齐/真值脚注与30–40FPS；真实记录数据评测。" } ], "corrections": [ { "field": "experimentType", "value": "data", "reason": "本文量化证据主要是真实传感器离线序列，不是机器人实机闭环任务。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04" }, "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-2007.00258", "title": "LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping", "titleZh": "LIO-SAM：通过平滑与建图实现紧耦合激光惯性里程计", "date": "2020-07-01", "datePrecision": "day", "year": 2020, "url": "https://arxiv.org/abs/2007.00258", "paperUrl": "https://arxiv.org/abs/2007.00258", "codeUrl": "https://github.com/TixiaoShan/LIO-SAM", "summary": "LIO-SAM以因子图紧耦合IMU预积分和激光里程计，可加入GPS和回环约束；IMU辅助去畸变并提供扫描配准初值，激光几何优化反过来修正偏置。按位移/转角选择关键帧，用局部子地图限制配准成本，iSAM2增量优化…", "abstractZh": "LIO-SAM以因子图紧耦合IMU预积分和激光里程计，可加入GPS和回环约束；IMU辅助去畸变并提供扫描配准初值，激光几何优化反过来修正偏置。按位移/转角选择关键帧，用局部子地图限制配准成本，iSAM2增量优化全局状态。\n五组真实传感器序列涵盖手持旋转/步行/校园、Clearpath Jackal林间和阿姆斯特丹船行；传感器为VLP-16、3DM-GX5-25与Reach M GPS。对比LOAM、LIOM，以及禁用GPS/回环的自身消融，评测漂移、地图一致性和速度。", "category": "激光惯性 / 因子图", "tags": [ "LIO-SAM", "LiDAR-IMU", "因子图", "预积分" ], "directions": [ "导航与建图" ], "tier": "classic", "experimentType": "data", "robots": [ "Clearpath Jackal（传感器采集平台）" ], "robotFilters": [ "Clearpath Jackal" ], "codeStatus": "open", "status": "实现已开源", "trainingNote": "无需神经网络训练；官方仓库提供 ROS 实现、样例数据和运行入口。需按 README 提供点级时间、ring 与正确 IMU 外参。", "contribution": "LIO-SAM以因子图紧耦合IMU预积分和激光里程计，可加入GPS和回环约束；IMU辅助去畸变并提供扫描配准初值，激光几何优化反过来修正偏置。按位移/转角选择关键帧，用局部子地图限制配准成本，iSAM2增量优化全局状态。", "whyUseful": "先核对激光每点时间戳、IMU坐标和外参，再分别运行仅LIO、加GPS、加回环。报告完整轨迹误差、末端漂移和每扫描耗时，避免只用回环后首尾误差宣称厘米级全程定位。", "limitations": "主要是记录数据上的定位建图验证，没有自主导航任务闭环成功率；GPS被作为评测参考又可参与估计，需明确不是独立完美真值。弱结构、校准和时间同步仍会影响，回环成立及绝对测量可用性不能省略。", "license": "BSD-3-Clause", "evidence": [ { "url": "https://arxiv.org/abs/2007.00258", "note": "首版日期、因子图架构与三个真实平台数据的实验范围。" }, { "url": "https://github.com/TixiaoShan/LIO-SAM", "note": "官方 BSD-3-Clause 源码；README 明确传感器格式、IMU、回环示例和固态雷达限制。" }, { "url": "https://arxiv.org/pdf/2007.00258", "note": "III; IV Table I：四类因子、传感器型号及五类采集平台。" }, { "url": "https://arxiv.org/pdf/2007.00258", "note": "Tables II–IV; V：末端漂移、GPS参考RMSE与实时范围。" }, { "url": "https://arxiv.org/pdf/2007.00258v3", "note": "III; IV Table I：四类因子、传感器型号及五类采集平台。" }, { "url": "https://arxiv.org/pdf/2007.00258v3", "note": "Tables II–IV; V：末端漂移、GPS参考RMSE与实时范围。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "用因子图把激光惯性里程计、GPS 与回环放到统一可扩展框架。", "experimentNote": "五组真实传感器序列涵盖手持旋转/步行/校园、Clearpath Jackal林间和阿姆斯特丹船行；传感器为VLP-16、3DM-GX5-25与Reach M GPS。对比LOAM、LIOM，以及禁用GPS/回环的自身消融，评测漂移、地图一致性和速度。", "projectUrl": null, "shortTitle": "", "original": { "id": "arxiv-2007.00258", "originalSourceUrl": "https://arxiv.org/pdf/2007.00258v3", "versionedOriginalUrl": "https://arxiv.org/pdf/2007.00258v3", "finalDownloadUrl": "https://arxiv.org/pdf/2007.00258", "sha256": "8ad11438fc0849de88b1604932f8bfd2732aee640aaf5fb41974b34f555c86f6", "bytes": 4368703, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:40:07.589827+00:00", "publicationDate": "2020/07/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2007.00258", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2007.00258", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2007.00258", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping", "sourceVersion": "2007.00258v3", "sourceVersionDate": "2020/07/14", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.383002+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2007.00258" }, "originalAnalysis": { "id": "arxiv-2007.00258", "sourceUrl": "https://arxiv.org/pdf/2007.00258v3", "sectionsRead": [ "III-A–E", "IV-A–E", "V", "Tables I–IV" ], "methodsZh": "LIO-SAM以因子图紧耦合IMU预积分和激光里程计，可加入GPS和回环约束；IMU辅助去畸变并提供扫描配准初值，激光几何优化反过来修正偏置。按位移/转角选择关键帧，用局部子地图限制配准成本，iSAM2增量优化全局状态。", "experimentsZh": "五组真实传感器序列涵盖手持旋转/步行/校园、Clearpath Jackal林间和阿姆斯特丹船行；传感器为VLP-16、3DM-GX5-25与Reach M GPS。对比LOAM、LIOM，以及禁用GPS/回环的自身消融，评测漂移、地图一致性和速度。", "resultsZh": "校园/公园/船行返回起点误差0.12/0.04/0.17米；公园对GPS轨迹RMSE0.96米，远低于无绝对约束的23.96米。加入回环明显修正GPS稀疏时的累积漂移，但“首尾接近”与全轨迹真实精度不是同一指标。", "limitationsZh": "主要是记录数据上的定位建图验证，没有自主导航任务闭环成功率；GPS被作为评测参考又可参与估计，需明确不是独立完美真值。弱结构、校准和时间同步仍会影响，回环成立及绝对测量可用性不能省略。", "reproductionZh": "先核对激光每点时间戳、IMU坐标和外参，再分别运行仅LIO、加GPS、加回环。报告完整轨迹误差、末端漂移和每扫描耗时，避免只用回环后首尾误差宣称厘米级全程定位。", "experimentType": "data", "robots": [ "Clearpath Jackal（传感器采集平台）" ], "corrections": [ "真实Jackal载传感器数据评测不是机器人自主闭环导航结果。" ], "evidenceNotes": [ { "note": "四类因子、传感器型号及五类采集平台。", "section": "III; IV Table I" }, { "note": "末端漂移、GPS参考RMSE与实时范围。", "section": "Tables II–IV; V" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "2007.00258v3", "originalSha256": "8ad11438fc0849de88b1604932f8bfd2732aee640aaf5fb41974b34f555c86f6" }, "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-2007.00643", "title": "Object Goal Navigation using Goal-Oriented Semantic Exploration", "titleZh": "利用目标导向语义探索进行物体目标导航", "date": "2020-07-01", "datePrecision": "day", "year": 2020, "url": "https://arxiv.org/abs/2007.00643", "paperUrl": "https://arxiv.org/abs/2007.00643", "codeUrl": "https://github.com/devendrachaplot/Object-Goal-Navigation", "summary": "SemExp以预训练Mask R-CNN识别RGB中的语义，再将深度点投到五厘米二维地图，学习去噪并随时间融合。高层PPO依据物体类别、语义地图和访问历史每25步选探索目标，找到目标类别后直接趋近；局部Fast…", "abstractZh": "SemExp以预训练Mask R-CNN识别RGB中的语义，再将深度点投到五厘米二维地图，学习去噪并随时间融合。高层PPO依据物体类别、语义地图和访问历史每25步选探索目标，找到目标类别后直接趋近；局部Fast Marching规划器每步重规划，分离感知、语义先验与几何控制。\n在Gibson和Matterport3D共86训练场景训练1000万帧，分别用1000/2000测试回合比较端到端RL、前沿探索与Active Neural SLAM。仿真使用完美深度和位姿、六类目标，500步内距目标一米停止为成功；竞赛另用21类和DeepLabv3，协议应区分。", "category": "具身导航 / 语义探索", "tags": [ "ObjectNav", "语义地图", "SemExp", "模块化策略" ], "directions": [ "导航与建图" ], "tier": "classic", "experimentType": "both", "robots": [ "LoCoBot" ], "robotFilters": [ "LoCoBot" ], "codeStatus": "open", "status": "实现已开源", "trainingNote": "官方 MIT 仓库提供 PyTorch 实现、预训练模型、训练与评测说明；需对应 Habitat/场景数据和语义感知依赖。", "contribution": "SemExp以预训练Mask R-CNN识别RGB中的语义，再将深度点投到五厘米二维地图，学习去噪并随时间融合。高层PPO依据物体类别、语义地图和访问历史每25步选探索目标，找到目标类别后直接趋近；局部Fast Marching规划器每步重规划，分离感知、语义先验与几何控制。", "whyUseful": "复现应保留地图类别通道、语义预训练来源、PPO奖励、86并行场景及场景划分，分别报Success、SPL和目标剩余距离。实机还需深度/里程计对齐及碰撞处理；竞赛与论文六类配置不能共用数字。", "limitations": "实机样本较少，未充分覆盖布局、动力学和目标类别变化；仿真完美感知使真实鲁棒性不能仅由仿真表推断。仍有约27%真值语义回合失败，主要未找到目标；目标阈值一米也不代表精细物体接近或抓取能力。", "license": "MIT", "evidence": [ { "url": "https://arxiv.org/abs/2007.00643", "note": "首版日期与目标导向语义地图方法。" }, { "url": "https://arxiv.org/pdf/2007.00643", "note": "§5.3 明确 LoCoBot 硬件和 PyRobot 实机迁移；实验包含仿真。" }, { "url": "https://devendrachaplot.github.io/projects/semantic-exploration", "note": "作者项目页说明语义建图、长期目标与解析局部规划模块。" }, { "url": "https://github.com/devendrachaplot/Object-Goal-Navigation", "note": "项目页直链官方实现，MIT 许可证及训练评测入口。" }, { "url": "https://arxiv.org/pdf/2007.00643v2", "note": "3/4：5cm cells;goal every25 steps;perfect simulator depth/pose;stop within1m;500-step horizon." }, { "url": "https://arxiv.org/pdf/2007.00643v2", "note": "5 表1/2：1000Gibson+2000MP3D episodes;54.4/36.0% success;GT semantics73.1%." }, { "url": "https://arxiv.org/pdf/2007.00643v2", "note": "5.2 表3：Challenge21 categories uses DeepLabv3,success25.3%,distinct from six-category paper setup." }, { "url": "https://arxiv.org/pdf/2007.00643v2", "note": "5.3：Locobot hardware+PyRobot;13/20 successful real-world trials." } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "把语义空间先验用于目标导向探索，展示学习策略与地图规划结合的导航路线。", "experimentNote": "在Gibson和Matterport3D共86训练场景训练1000万帧，分别用1000/2000测试回合比较端到端RL、前沿探索与Active Neural SLAM。仿真使用完美深度和位姿、六类目标，500步内距目标一米停止为成功；竞赛另用21类和DeepLabv3，协议应区分。", "projectUrl": null, "shortTitle": "", "original": { "id": "arxiv-2007.00643", "originalSourceUrl": "https://arxiv.org/pdf/2007.00643v2", "versionedOriginalUrl": "https://arxiv.org/pdf/2007.00643v2", "finalDownloadUrl": "https://arxiv.org/pdf/2007.00643", "sha256": "abe66c10e6d8115852edf797cc02bd09b15e587e1690573c4582cf7d714621f2", "bytes": 1965159, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:40:12.576230+00:00", "publicationDate": "2020/07/01", "publicationDateSourceUrl": "https://arxiv.org/abs/2007.00643", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2007.00643", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2007.00643", "pages": 11, "sourceIdentityStatus": "title_match", "sourceTitle": "Object Goal Navigation using Goal-Oriented Semantic Exploration", "sourceVersion": "2007.00643v2", "sourceVersionDate": "2020/07/02", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.385579+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2007.00643" }, "originalAnalysis": { "id": "arxiv-2007.00643", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739408+00:00", "sourceUrl": "https://arxiv.org/pdf/2007.00643v2", "sectionsRead": [ "3 语义地图、高层目标与局部规划", "4 数据集、传感器、PPO和指标", "5 表1–3、消融与错误分析", "5.3 LoCoBot真实迁移" ], "methodsZh": "SemExp以预训练Mask R-CNN识别RGB中的语义，再将深度点投到五厘米二维地图，学习去噪并随时间融合。高层PPO依据物体类别、语义地图和访问历史每25步选探索目标，找到目标类别后直接趋近；局部Fast Marching规划器每步重规划，分离感知、语义先验与几何控制。", "experimentsZh": "在Gibson和Matterport3D共86训练场景训练1000万帧，分别用1000/2000测试回合比较端到端RL、前沿探索与Active Neural SLAM。仿真使用完美深度和位姿、六类目标，500步内距目标一米停止为成功；竞赛另用21类和DeepLabv3，协议应区分。", "resultsZh": "作者报告Gibson/MP3D成功率54.4%/36.0%，SPL为0.199/0.144，高于所比探索基线。真值语义把Gibson成功率提高到73.1%，揭示感知误差瓶颈。真实LoCoBot用PyRobot部署，20次试验成功13次即65%，支持有限真实迁移。", "limitationsZh": "实机样本较少，未充分覆盖布局、动力学和目标类别变化；仿真完美感知使真实鲁棒性不能仅由仿真表推断。仍有约27%真值语义回合失败，主要未找到目标；目标阈值一米也不代表精细物体接近或抓取能力。", "reproductionZh": "复现应保留地图类别通道、语义预训练来源、PPO奖励、86并行场景及场景划分，分别报Success、SPL和目标剩余距离。实机还需深度/里程计对齐及碰撞处理；竞赛与论文六类配置不能共用数字。", "evidenceNotes": [ { "section/page": "3/4", "note": "5cm cells;goal every25 steps;perfect simulator depth/pose;stop within1m;500-step horizon." }, { "section/page": "5 表1/2", "note": "1000Gibson+2000MP3D episodes;54.4/36.0% success;GT semantics73.1%." }, { "section/page": "5.2 表3", "note": "Challenge21 categories uses DeepLabv3,success25.3%,distinct from six-category paper setup." }, { "section/page": "5.3", "note": "Locobot hardware+PyRobot;13/20 successful real-world trials." } ], "experimentType": "both", "robots": [ "LoCoBot" ], "corrections": [ "65%只对应20次实机试验；与六类仿真及21类竞赛成绩分列。" ], "sourceVersion": "2007.00643v2", "originalSha256": "abe66c10e6d8115852edf797cc02bd09b15e587e1690573c4582cf7d714621f2" }, "analysisVerifiedAt": "2026-10-04T13:47:58.739408+00:00" }, { "id": "arxiv-1912.01603", "title": "Dream to Control: Learning Behaviors by Latent Imagination", "titleZh": "想象以控制：通过潜在想象学习行为", "shortTitle": "Dreamer", "date": "2019-12-03", "year": 2019, "url": "https://arxiv.org/abs/1912.01603", "paperUrl": "https://arxiv.org/abs/1912.01603", "projectUrl": "https://dreamrl.github.io/", "codeUrl": "https://github.com/danijar/dreamer", "summary": "Dreamer先学习包含确定性记忆与随机潜变量的世界模型，再从真实回放的潜状态出发想象短轨迹，用价值模型补足规划窗外回报。连续动作通过可微动力学反传解析梯度优化actor，部署执行策略网络，无需每步像PlaNe…", "abstractZh": "Dreamer先学习包含确定性记忆与随机潜变量的世界模型，再从真实回放的潜状态出发想象短轨迹，用价值模型补足规划窗外回报。连续动作通过可微动力学反传解析梯度优化actor，部署执行策略网络，无需每步像PlaNet一样搜索动作序列。\n主评估为DeepMind Control Suite二十视觉任务，64×64RGB、1–12维动作、1000步回合、动作重复2和随机初态。对照重跑的PlaNet及文献D4PG/A3C，另测Atari/DeepMind Lab的离散动作和提前终止，消融价值与表征学习目标。", "category": "世界模型 / 想象强化学习", "tags": [ "Dreamer", "世界模型", "强化学习" ], "directions": [ "世界模型", "强化学习" ], "tier": "foundation", "experimentType": "sim", "experimentNote": "主评估为DeepMind Control Suite二十视觉任务，64×64RGB、1–12维动作、1000步回合、动作重复2和随机初态。对照重跑的PlaNet及文献D4PG/A3C，另测Atari/DeepMind Lab的离散动作和提前终止，消融价值与表征学习目标。", "robots": [], "robotFilters": [], "robotNote": "无实机；仿真形态不等同于已验证的商业机器人型号。", "codeStatus": "open", "status": "训练代码已开源", "trainingStatus": "训练代码已开源", "trainingNote": "作者 TensorFlow 2 实现含训练流程和分数，MIT 许可；README 另指向论文原始 TensorFlow 1 实现。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "MIT（所列 TF2 代码）", "contribution": "Dreamer先学习包含确定性记忆与随机潜变量的世界模型，再从真实回放的潜状态出发想象短轨迹，用价值模型补足规划窗外回报。连续动作通过可微动力学反传解析梯度优化actor，部署执行策略网络，无需每步像PlaNet一样搜索动作序列。", "whyUseful": "复现需明确环境步与动作重复计数、随机初态、相同PlaNet重复率和归一化得分，分别记录actor/value/world model更新。正文提供源码与超参数，但本轮未执行；离散实验是适用性补充，不等于后续DreamerV2全Atari成绩。", "limitations": "所有证据为模拟控制/游戏，没有实体机器人。模型误差可能影响长时域想象，成功依赖可学习的视觉表征；跨算法训练预算与硬件来源不同，不能把步数优势直接换算成任何系统的墙钟加速。", "caveats": "主要是仿真视觉控制；策略收益受奖励定义及世界模型误差影响。", "evidence": [ { "url": "https://arxiv.org/abs/1912.01603", "note": "首发日期。" }, { "url": "https://arxiv.org/html/1912.01603v3", "note": "原文算法、潜在想象与视觉控制评估。" }, { "url": "https://github.com/danijar/dreamer", "note": "作者 TF2 实现、MIT 及原始实现区别。" }, { "url": "https://arxiv.org/pdf/1912.01603", "note": "§6：20视觉任务；500万步均823，D4PG 1亿步786。" }, { "url": "https://arxiv.org/pdf/1912.01603", "note": "§3 / Fig.4：价值估计补足有限想象窗口。" }, { "url": "https://arxiv.org/pdf/1912.01603", "note": "Appendix C：离散动作/提前结束为补充模拟实验，无实机。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2019：从模型内动作搜索推进到用想象轨迹学习可直接执行的策略。", "freshness": "基础奠基", "original": { "id": "arxiv-1912.01603", "originalSourceUrl": "https://arxiv.org/pdf/1912.01603v3", "versionedOriginalUrl": "https://arxiv.org/pdf/1912.01603v3", "finalDownloadUrl": "https://arxiv.org/pdf/1912.01603", "sha256": "318891d59ba0065f57eb0e735735700e5eb528f53bf08b81c773437385586c64", "bytes": 1999946, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:40:12.645695+00:00", "publicationDate": "2019/12/03", "publicationDateSourceUrl": "https://arxiv.org/abs/1912.01603", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1912.01603", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1912.01603", "pages": 20, "sourceIdentityStatus": "title_match", "sourceTitle": "Dream to Control: Learning Behaviors by Latent Imagination", "sourceVersion": "1912.01603v3", "sourceVersionDate": "2020/03/17", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.388006+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1912.01603" }, "originalAnalysis": { "id": "arxiv-1912.01603", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090148+00:00", "sourceUrl": "https://arxiv.org/pdf/1912.01603", "sectionsRead": [ "§3–4", "§6", "Figs.4,6,8", "§7", "Appendix A,C" ], "methodsZh": "Dreamer先学习包含确定性记忆与随机潜变量的世界模型，再从真实回放的潜状态出发想象短轨迹，用价值模型补足规划窗外回报。连续动作通过可微动力学反传解析梯度优化actor，部署执行策略网络，无需每步像PlaNet一样搜索动作序列。", "experimentsZh": "主评估为DeepMind Control Suite二十视觉任务，64×64RGB、1–12维动作、1000步回合、动作重复2和随机初态。对照重跑的PlaNet及文献D4PG/A3C，另测Atari/DeepMind Lab的离散动作和提前终止，消融价值与表征学习目标。", "resultsZh": "作者报告500万环境步后任务平均823，D4PG在1亿步内平均786；每100万步单V100约三小时，PlaNet约十一小时。价值自举降低对想象窗口长度的敏感性，图像重建在该组任务优于仅预测奖励。", "limitationsZh": "所有证据为模拟控制/游戏，没有实体机器人。模型误差可能影响长时域想象，成功依赖可学习的视觉表征；跨算法训练预算与硬件来源不同，不能把步数优势直接换算成任何系统的墙钟加速。", "reproductionZh": "复现需明确环境步与动作重复计数、随机初态、相同PlaNet重复率和归一化得分，分别记录actor/value/world model更新。正文提供源码与超参数，但本轮未执行；离散实验是适用性补充，不等于后续DreamerV2全Atari成绩。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "§6", "note": "20视觉任务；500万步均823，D4PG 1亿步786。" }, { "section/page": "§3 / Fig.4", "note": "价值估计补足有限想象窗口。" }, { "section/page": "Appendix C", "note": "离散动作/提前结束为补充模拟实验，无实机。" } ] }, "analysisVerifiedAt": "2026-10-04T13:50:35.090148+00:00" }, { "id": "arxiv-1910.10897", "title": "Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning", "date": "2019-10-24", "url": "https://arxiv.org/abs/1910.10897v2", "titleZh": "Meta-World：多任务与元强化学习的机器人操作基准", "abstractZh": "设计50类共享控制结构但物体和交互不同的桌面任务，区分同任务目标变化、多任务学习以及未见任务的少样本适应。\n提出MT10、MT50、ML1、ML10、ML45等设置，比较PPO、TRPO、SAC、MAML、RL²和PEARL；成功由任务相关距离阈值定义。", "summary": "设计50类共享控制结构但物体和交互不同的桌面任务，区分同任务目标变化、多任务学习以及未见任务的少样本适应。", "experimentType": "sim", "robots": [ "Rethink Sawyer（仿真）" ], "tags": [ "数据集与基准", "强化学习", "操作与抓取" ], "limitations": "只验证仿真桌面Sawyer；任务共享动力学及场景结构，不能代表任意机器人和现实开放世界泛化。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "foundation", "tierNote": "编辑分类：多任务及元强化学习的基础评估框架，MT/ML任务拆分是理解后续比较的重要前提。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "官方实现已核验", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/1910.10897v2", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/1910.10897v2", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/Farama-Foundation/Metaworld", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "open", "codeUrl": "https://github.com/Farama-Foundation/Metaworld", "category": "数据集与基准 / 强化学习", "year": 2019, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "环境、专家策略可用；README 另链论文算法仓库。 脚本专家可生成模仿学习示范。 基准本身不依赖模型权重。 当前 Meta-World+ 与原论文旧版有差异；比较结果必须固定环境版本、奖励、任务划分和终止规则。 本次未运行训练、复现实验或实机控制。", "directions": [ "数据集与基准", "强化学习", "操作与抓取" ], "robotFilters": [ "Rethink Sawyer（仿真）" ], "original": { "id": "arxiv-1910.10897", "originalSourceUrl": "https://arxiv.org/pdf/1910.10897v2", "versionedOriginalUrl": "https://arxiv.org/pdf/1910.10897v2", "finalDownloadUrl": "https://arxiv.org/pdf/1910.10897v2", "sha256": "8ea4fd9e7a2f7a2cfc475ae7e7021e7a6c03b99ae3932b455def9197ff1197e8", "bytes": 6237424, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T14:59:57.241402+00:00", "publicationDate": "2019/10/24", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1910.10897v2", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1910.10897v2", "pages": 40, "sourceIdentityStatus": "title_match", "sourceTitle": "Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning", "sourceVersion": "1910.10897v2", "sourceVersionDate": "2021/06/14", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/1910.10897v2", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1910.10897" }, "originalAnalysis": { "id": "arxiv-1910.10897", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/1910.10897v2", "sourceVersion": "1910.10897v2", "sectionsRead": [ "PDF第2页选段（提取文本L77–90）", "PDF第4页选段（提取文本L176–193）", "PDF第7页选段（提取文本L333–350）", "PDF第7页选段（提取文本L352–365）", "PDF第6–7页选段（提取文本L282–325）", "PDF第12页选段（提取文本L730–752）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "设计50类共享控制结构但物体和交互不同的桌面任务，区分同任务目标变化、多任务学习以及未见任务的少样本适应。", "experimentsZh": "提出MT10、MT50、ML1、ML10、ML45等设置，比较PPO、TRPO、SAC、MAML、RL²和PEARL；成功由任务相关距离阈值定义。", "resultsZh": "文中MT-SAC在MT10约68%成功，但扩展到50类时明显下降；当时的元学习方法连多样训练任务也难充分学会。", "limitationsZh": "只验证仿真桌面Sawyer；任务共享动力学及场景结构，不能代表任意机器人和现实开放世界泛化。", "reproductionZh": "锁定基准版本、任务拆分、目标是否进入观测、奖励与成功阈值；不能把MT训练任务成功与ML未见任务适应混为一谈。", "experimentType": "sim", "robots": [ "Rethink Sawyer（仿真）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "设计50类共享控制结构但物体和交互不同的桌面任务，区分同任务目标变化、多任务学习以及未见任务的少样本适应。" }, { "section": "原文实验与结果相关选段", "note": "提出MT10、MT50、ML1、ML10、ML45等设置，比较PPO、TRPO、SAC、MAML、RL²和PEARL；成功由任务相关距离阈值定义。 文中MT-SAC在MT10约68%成功，但扩展到50类时明显下降；当时的元学习方法连多样训练任务也难充分学会。" }, { "section": "局限与复现条件", "note": "只验证仿真桌面Sawyer；任务共享动力学及场景结构，不能代表任意机器人和现实开放世界泛化。" } ], "originalSha256": "8ea4fd9e7a2f7a2cfc475ae7e7021e7a6c03b99ae3932b455def9197ff1197e8", "readingScopeSha256": "ddf5571219b7f0cf15333697a11b35271b9f9d004867d4ad587501068d5bbf7c", "textExtractionSha256": "ec09de483958e349b355822660eb1a173564084793da7746bd4128fee31c4845" }, "experimentNote": "提出MT10、MT50、ML1、ML10、ML45等设置，比较PPO、TRPO、SAC、MAML、RL²和PEARL；成功由任务相关距离阈值定义。", "contribution": "设计50类共享控制结构但物体和交互不同的桌面任务，区分同任务目标变化、多任务学习以及未见任务的少样本适应。", "whyUseful": "锁定基准版本、任务拆分、目标是否进入观测、奖励与成功阈值；不能把MT训练任务成功与ML未见任务适应混为一谈。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "仓库 MIT；外部算法实现和附加数据各自核对。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-1910.07113", "title": "Solving Rubik's Cube with a Robot Hand", "titleZh": "用机器人手解魔方", "shortTitle": "Rubik’s Cube / ADR", "date": "2019-10-16", "datePrecision": "day", "year": 2019, "url": "https://arxiv.org/abs/1910.07113", "paperUrl": "https://arxiv.org/abs/1910.07113", "codeUrl": "https://github.com/openai/robogym", "summary": "自动域随机化根据边界环境成绩逐步扩展质量、摩擦、视觉等参数范围，在模拟中训练循环控制策略和视觉状态估计器，再迁移到Shadow灵巧手。Kociemba算法提供魔方解的子目标序列，学习策略负责翻转整块和转动指定面…", "abstractZh": "自动域随机化根据边界环境成绩逐步扩展质量、摩擦、视觉等参数范围，在模拟中训练循环控制策略和视觉状态估计器，再迁移到Shadow灵巧手。Kociemba算法提供魔方解的子目标序列，学习策略负责翻转整块和转动指定面，而非自行搜索符号解法。\n先以方块重定向检验随机化迁移，再每策略十次真实魔方试验。量化实验固定一条随机公平打乱序列，从已解状态执行，需26次面转与17次翻转；比较人工随机化、不同规模ADR以及Giiker面角传感与纯视觉。", "category": "灵巧手 / 自动域随机化", "tags": [ "ADR", "手内操作", "循环策略", "Sim2Real" ], "directions": [ "灵巧手", "操作与抓取" ], "tier": "classic", "experimentType": "both", "robots": [ "Shadow Dexterous Hand" ], "robotNote": "", "codeStatus": "open", "status": "官方仿真环境公开；不等于完整分布式训练与硬件复现包。", "trainingNote": "robogym提供Dactyl及魔方环境和随机化接口；本次未核实论文全部ADR、视觉训练和部署流程可一键复现。", "contribution": "自动域随机化根据边界环境成绩逐步扩展质量、摩擦、视觉等参数范围，在模拟中训练循环控制策略和视觉状态估计器，再迁移到Shadow灵巧手。Kociemba算法提供魔方解的子目标序列，学习策略负责翻转整块和转动指定面，而非自行搜索符号解法。", "whyUseful": "需区分魔方面角传感器、相机位姿、符号求解器和控制器，保留ADR范围扩展、子动作超时、落物终止及训练预算。重现视频与表6必须说明各自初态和目标协议，本次未运行系统。", "limitations": "最优策略使用数月大规模训练和专门硬件，十次试验且固定序列不能证明任意打乱稳定解决。循环状态适应与物理参数辨识有关，但不等于部署中更新网络权重；完全视觉结果仍明显较弱。", "evidence": [ { "url": "https://arxiv.org/abs/1910.07113", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/pdf/1910.07113", "note": "§2说明Kociemba子目标；§3.1明确Shadow E3M5R与硬件改装。" }, { "url": "https://github.com/openai/robogym", "note": "官方README列出Shadow手的Dactyl锁定方块和魔方环境。" }, { "url": "https://arxiv.org/pdf/1910.07113", "note": "§2; §5–7：Kociemba提供序列，ADR用于视觉与循环控制训练。" }, { "url": "https://arxiv.org/pdf/1910.07113", "note": "§8.4.1 Table 6; §8.4.2：固定序列十次、20%需Giiker、纯视觉0/10，扰动未量化。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2019：从手内重定向推进到高精度多阶段操作，并把域随机化自动化。", "projectUrl": null, "robotFilters": [ "Shadow Hand" ], "original": { "id": "arxiv-1910.07113", "originalSourceUrl": "https://arxiv.org/pdf/1910.07113v1", "versionedOriginalUrl": "https://arxiv.org/pdf/1910.07113v1", "finalDownloadUrl": "https://arxiv.org/pdf/1910.07113", "sha256": "2a6fe43013ddc8d148b0cdebc8468dfe83d11319ed626a1aeaa79b510e2ed48c", "bytes": 5194320, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:40:21.511874+00:00", "publicationDate": "2019/10/16", "publicationDateSourceUrl": "https://arxiv.org/abs/1910.07113", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1910.07113", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1910.07113", "pages": 51, "sourceIdentityStatus": "title_match", "sourceTitle": "Solving Rubik's Cube with a Robot Hand", "sourceVersion": "1910.07113v1", "sourceVersionDate": "2019/10/16", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.392905+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1910.07113" }, "originalAnalysis": { "id": "arxiv-1910.07113", "sourceUrl": "https://arxiv.org/pdf/1910.07113", "sectionsRead": [ "§2–3", "§5–7", "§8.1–8.4", "Table 6", "§9,11" ], "methodsZh": "自动域随机化根据边界环境成绩逐步扩展质量、摩擦、视觉等参数范围，在模拟中训练循环控制策略和视觉状态估计器，再迁移到Shadow灵巧手。Kociemba算法提供魔方解的子目标序列，学习策略负责翻转整块和转动指定面，而非自行搜索符号解法。", "experimentsZh": "先以方块重定向检验随机化迁移，再每策略十次真实魔方试验。量化实验固定一条随机公平打乱序列，从已解状态执行，需26次面转与17次翻转；比较人工随机化、不同规模ADR以及Giiker面角传感与纯视觉。", "resultsZh": "作者报告最大ADR策略在视觉位姿＋Giiker面角条件下平均26.8次成功子动作，半序列60%、全序列20%；纯视觉面角时全序列0/10。视频有从打乱到复原的展示，但不能用它替代固定协议量化；手套、遮挡等扰动仅定性。", "limitationsZh": "最优策略使用数月大规模训练和专门硬件，十次试验且固定序列不能证明任意打乱稳定解决。循环状态适应与物理参数辨识有关，但不等于部署中更新网络权重；完全视觉结果仍明显较弱。", "reproductionZh": "需区分魔方面角传感器、相机位姿、符号求解器和控制器，保留ADR范围扩展、子动作超时、落物终止及训练预算。重现视频与表6必须说明各自初态和目标协议，本次未运行系统。", "experimentType": "both", "robots": [ "Shadow Dexterous Hand" ], "evidenceNotes": [ { "section/page": "§2; §5–7", "note": "Kociemba提供序列，ADR用于视觉与循环控制训练。" }, { "section/page": "§8.4.1 Table 6; §8.4.2", "note": "固定序列十次、20%需Giiker、纯视觉0/10，扰动未量化。" } ], "corrections": [ { "field": "resultsNote", "value": "完整公平序列20%为视觉位姿＋Giiker面角；纯视觉面角0/10", "reason": "Table 6区分传感配置，不能统称纯视觉20%。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04" }, "experimentNote": "先以方块重定向检验随机化迁移，再每策略十次真实魔方试验。量化实验固定一条随机公平打乱序列，从已解状态执行，需26次面转与17次翻转；比较人工随机化、不同规模ADR以及Giiker面角传感与纯视觉。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-1910.03135", "title": "DexPilot: Vision Based Teleoperation of Dexterous Robotic Hand-Arm System", "titleZh": "DexPilot：基于视觉的灵巧机器人手臂遥操作系统", "shortTitle": "DexPilot", "date": "2019-10-07", "datePrecision": "day", "year": 2019, "url": "https://arxiv.org/abs/1910.03135", "paperUrl": "https://arxiv.org/abs/1910.03135", "codeUrl": null, "summary": "DexPilot以多相机深度手部追踪和神经关键点估计驱动灵巧遥操作，将人手指尖距离/方向而非关节角直接匹配到Allegro。非线性重定向在指尖接近时做接触投影，结合RMP机械臂运动生成及低层关节控制，使四指机器…", "abstractZh": "DexPilot以多相机深度手部追踪和神经关键点估计驱动灵巧遥操作，将人手指尖距离/方向而非关节角直接匹配到Allegro。非线性重定向在指尖接近时做接触投影，结合RMP机械臂运动生成及低层关节控制，使四指机器人近似人手精细抓捏。\n真实KUKA LBR iiwa7 R800与Wonik Allegro共23自由度；两名熟悉操作者先暖身再各连续五次，15类任务含抓放、三尺寸堆块、倒珠、开罐、手内转砖、取钱、滑卡及抽屉操作。记录成功和完成时间，掉出工作区计失败。", "category": "灵巧手 / 动作重定向", "tags": [ "遥操作", "手部重定向", "示范采集", "多指操作" ], "directions": [ "灵巧手", "操作与抓取", "模仿学习" ], "tier": "classic", "experimentType": "real", "robots": [ "KUKA LBR iiwa7 R800", "Wonik Robotics Allegro Hand" ], "robotNote": "", "codeStatus": "unknown", "status": "作者项目展示公开；本次未核实完整官方代码。", "trainingNote": "核心是视觉追踪、重定向优化与底层控制；自主策略训练被作为未来用途讨论。", "contribution": "DexPilot以多相机深度手部追踪和神经关键点估计驱动灵巧遥操作，将人手指尖距离/方向而非关节角直接匹配到Allegro。非线性重定向在指尖接近时做接触投影，结合RMP机械臂运动生成及低层关节控制，使四指机器人近似人手精细抓捏。", "whyUseful": "先验证相机同步、手/机器人基准坐标、指尖距离阈值和安全碰撞平面，再按连续试验协议测不同操作者。应公开失败和重抓时间，另行训练策略后才能宣称学习收益；触觉记录不等于触觉反馈。", "limitations": "这是有人遥操作能力，尚非自主策略训练/部署结果；没有操作者触觉反馈，手部遮挡、接触识别和投影误差会妨碍精确控制。仅两操作者及每项五次，不能推断广泛用户可用性或自主泛化。", "evidence": [ { "url": "https://arxiv.org/abs/1910.03135", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/pdf/1910.03135", "note": "§III明确KUKA LBR iiwa7 R800与Allegro Hand；§VIII为真实遥操作实验。" }, { "url": "https://sites.google.com/view/dex-pilot", "note": "作者页面说明两名操作者、无触觉反馈以及示范用途。" }, { "url": "https://arxiv.org/pdf/1910.03135", "note": "III; VII：精确硬件、指尖几何重定向与RMP。" }, { "url": "https://arxiv.org/pdf/1910.03135", "note": "VIII Table II; IX–X：两操作者、连续五次协议、无触觉反馈及插装失败。" }, { "url": "https://arxiv.org/pdf/1910.03135v2", "note": "III; VII：精确硬件、指尖几何重定向与RMP。" }, { "url": "https://arxiv.org/pdf/1910.03135v2", "note": "VIII Table II; IX–X：两操作者、连续五次协议、无触觉反馈及插装失败。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2019：把视觉手部重定向接入真实灵巧操作，为后续人类示范学习铺路。", "projectUrl": null, "robotFilters": [ "KUKA LBR iiwa7 R800", "Allegro Hand" ], "original": { "id": "arxiv-1910.03135", "originalSourceUrl": "https://arxiv.org/pdf/1910.03135v2", "versionedOriginalUrl": "https://arxiv.org/pdf/1910.03135v2", "finalDownloadUrl": "https://arxiv.org/pdf/1910.03135", "sha256": "d04e6f165fdfd64b02de0545d69feef99ae77351c3635f9190394594aad472d1", "bytes": 5763327, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:40:21.531823+00:00", "publicationDate": "2019/10/07", "publicationDateSourceUrl": "https://arxiv.org/abs/1910.03135", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/1910.03135", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1910.03135", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "DexPilot: Vision Based Teleoperation of Dexterous Robotic Hand-Arm System", "sourceVersion": "1910.03135v2", "sourceVersionDate": "2019/10/14", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.401318+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1910.03135" }, "originalAnalysis": { "id": "arxiv-1910.03135", "sourceUrl": "https://arxiv.org/pdf/1910.03135v2", "sectionsRead": [ "III–VII", "VIII", "IX–X", "Table II; Figs 14–15" ], "methodsZh": "DexPilot以多相机深度手部追踪和神经关键点估计驱动灵巧遥操作，将人手指尖距离/方向而非关节角直接匹配到Allegro。非线性重定向在指尖接近时做接触投影，结合RMP机械臂运动生成及低层关节控制，使四指机器人近似人手精细抓捏。", "experimentsZh": "真实KUKA LBR iiwa7 R800与Wonik Allegro共23自由度；两名熟悉操作者先暖身再各连续五次，15类任务含抓放、三尺寸堆块、倒珠、开罐、手内转砖、取钱、滑卡及抽屉操作。记录成功和完成时间，掉出工作区计失败。", "resultsZh": "系统展示多个原先难以手工编程的多阶段灵巧任务，并能同步记录机器人状态、命令与触觉供未来示范学习；成功依赖操作者策略及重抓。额外高精度NIST插装整体不理想，特定条件成功率通常仅约10%。", "limitationsZh": "这是有人遥操作能力，尚非自主策略训练/部署结果；没有操作者触觉反馈，手部遮挡、接触识别和投影误差会妨碍精确控制。仅两操作者及每项五次，不能推断广泛用户可用性或自主泛化。", "reproductionZh": "先验证相机同步、手/机器人基准坐标、指尖距离阈值和安全碰撞平面，再按连续试验协议测不同操作者。应公开失败和重抓时间，另行训练策略后才能宣称学习收益；触觉记录不等于触觉反馈。", "experimentType": "real", "robots": [ "KUKA LBR iiwa7 R800", "Wonik Robotics Allegro Hand" ], "corrections": [ "原文是遥操作与数据采集系统，学习自主策略被列为未来方向。" ], "evidenceNotes": [ { "note": "精确硬件、指尖几何重定向与RMP。", "section": "III; VII" }, { "note": "两操作者、连续五次协议、无触觉反馈及插装失败。", "section": "VIII Table II; IX–X" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "1910.03135v2", "originalSha256": "d04e6f165fdfd64b02de0545d69feef99ae77351c3635f9190394594aad472d1" }, "experimentNote": "真实KUKA LBR iiwa7 R800与Wonik Allegro共23自由度；两名熟悉操作者先暖身再各连续五次，15类任务含抓放、三尺寸堆块、倒珠、开罐、手内转砖、取钱、滑卡及抽屉操作。记录成功和完成时间，掉出工作区计失败。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-1909.12271", "title": "RLBench: The Robot Learning Benchmark & Learning Environment", "date": "2019-09-26", "url": "https://arxiv.org/abs/1909.12271v1", "titleZh": "RLBench：机器人学习基准与学习环境", "abstractZh": "在统一V-REP/PyRep场景中封装任务变化、传感器、成功判据和示教生成器，支持多任务学习及少样本模仿。\n原版包含100个任务，meta-test保留10%；提出1、5、20条示教的未见任务评估，提供RGB、深度、分割与本体状态。", "summary": "在统一V-REP/PyRep场景中封装任务变化、传感器、成功判据和示教生成器，支持多任务学习及少样本模仿。", "experimentType": "sim", "robots": [ "Franka Emika Panda（仿真）" ], "tags": [ "数据集与基准", "模仿学习", "操作与抓取" ], "limitations": "统一仿真平台便于比较，但现实传感器和接触动力学仍有差距；任务集会扩展，跨版本分数不可直接比较。", "translationType": "中文原文选段分析（非逐字全文翻译）", "tier": "foundation", "tierNote": "编辑分类：机器人多任务与少样本模仿常用基准基础设施，提供标准任务、示教与评估入口。", "tierBasis": "编辑选读分类，非论文事实、作者自评或质量评级；不按年份自动归类。", "status": "源码可用，存在使用限制", "verification": "verified", "evidence": [ { "url": "https://arxiv.org/abs/1909.12271v1", "note": "原文标题、首发日期和版本已核验；PDF下载哈希已验证。" }, { "url": "https://arxiv.org/pdf/1909.12271v1", "note": "已读范围见原文分析的逐段页码/提取文本行号；方法、实验、结果、局限均以PDF选段核对。" }, { "url": "https://github.com/stepjam/RLBench", "note": "已发布实现，非占位仓库；代码、模型和数据条款应分开核实。" } ], "codeStatus": "source_available", "codeUrl": "https://github.com/stepjam/RLBench", "category": "数据集与基准 / 模仿学习", "year": 2019, "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "trainingNote": "环境和数据生成入口已发布；算法训练由使用者或上层项目提供。 可按任务程序化生成示范，非统一固定大模型数据集。 环境本身不需要预训练策略权重。 不能将源码公开误写为宽松开源；换机械臂可能导致任务不可解，正式基准应保留 Panda。 本次未运行训练、复现实验或实机控制。", "directions": [ "数据集与基准", "模仿学习", "操作与抓取" ], "robotFilters": [ "Franka Emika Panda（仿真）" ], "original": { "id": "arxiv-1909.12271", "originalSourceUrl": "https://arxiv.org/pdf/1909.12271v1", "versionedOriginalUrl": "https://arxiv.org/pdf/1909.12271v1", "finalDownloadUrl": "https://arxiv.org/pdf/1909.12271v1", "sha256": "fbb04509bea7bf6aa9b97b07c46d95e5c1a900987599693b3fae3fd84377db76", "bytes": 3732227, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T14:59:53.906074+00:00", "publicationDate": "2019/09/26", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1909.12271v1", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1909.12271v1", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "RLBench: The Robot Learning Benchmark & Learning Environment", "sourceVersion": "1909.12271v1", "sourceVersionDate": "2019/09/26", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T15:21:15.250223+00:00", "publicationDateSourceUrl": "https://arxiv.org/abs/1909.12271v1", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1909.12271" }, "originalAnalysis": { "id": "arxiv-1909.12271", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T15:21:15.250223+00:00", "sourceUrl": "https://arxiv.org/pdf/1909.12271v1", "sourceVersion": "1909.12271v1", "sectionsRead": [ "PDF第3页选段（提取文本L198–215）", "PDF第4页选段（提取文本L254–267）", "PDF第5页选段（提取文本L440–475）", "PDF第6页选段（提取文本L513–544）" ], "readingScope": "PDF原文选段分析：读取方法、实验/结果及局限相关段落；不是全篇逐段精读或逐字全文翻译。范围判断和复现提醒包含基于已读实验条件的分析。", "methodsZh": "在统一V-REP/PyRep场景中封装任务变化、传感器、成功判据和示教生成器，支持多任务学习及少样本模仿。", "experimentsZh": "原版包含100个任务，meta-test保留10%；提出1、5、20条示教的未见任务评估，提供RGB、深度、分割与本体状态。", "resultsZh": "贡献主要是可重复的任务与示教基础设施；文中任务长度约100至1000步，不能把基准发布当作某种策略达到高成功率。", "limitationsZh": "统一仿真平台便于比较，但现实传感器和接触动力学仍有差距；任务集会扩展，跨版本分数不可直接比较。", "reproductionZh": "锁定任务集版本、train/test任务拆分、变化编号、动作模式和CoppeliaSim/PyRep依赖，报告少样本设置。", "experimentType": "sim", "robots": [ "Franka Emika Panda（仿真）" ], "evidenceNotes": [ { "section": "原文方法相关选段", "note": "在统一V-REP/PyRep场景中封装任务变化、传感器、成功判据和示教生成器，支持多任务学习及少样本模仿。" }, { "section": "原文实验与结果相关选段", "note": "原版包含100个任务，meta-test保留10%；提出1、5、20条示教的未见任务评估，提供RGB、深度、分割与本体状态。 贡献主要是可重复的任务与示教基础设施；文中任务长度约100至1000步，不能把基准发布当作某种策略达到高成功率。" }, { "section": "局限与复现条件", "note": "统一仿真平台便于比较，但现实传感器和接触动力学仍有差距；任务集会扩展，跨版本分数不可直接比较。" } ], "originalSha256": "fbb04509bea7bf6aa9b97b07c46d95e5c1a900987599693b3fae3fd84377db76", "readingScopeSha256": "2653a563bbda77f4f98ad36a0dc3493906100eda8d015dac28b17d3beedb868f", "textExtractionSha256": "a3840f4515c7036e2b4d1ac2783c1e4b6f9f57d4ddb79b77b17e9fb39a5851f3" }, "experimentNote": "原版包含100个任务，meta-test保留10%；提出1、5、20条示教的未见任务评估，提供RGB、深度、分割与本体状态。", "contribution": "在统一V-REP/PyRep场景中封装任务变化、传感器、成功判据和示教生成器，支持多任务学习及少样本模仿。", "whyUseful": "锁定任务集版本、train/test任务拆分、变化编号、动作模式和CoppeliaSim/PyRep依赖，报告少样本设置。", "analysisVerifiedAt": "2026-10-04T15:21:15.250223+00:00", "codeLicenseNote": "RLBench 自定义许可仅允许非商业、内部或学术研究；CoppeliaSim 等依赖另有许可。", "codeReleaseScope": "已发布实现，非占位仓库" }, { "id": "arxiv-1904.01201", "title": "Habitat: A Platform for Embodied AI Research", "titleZh": "Habitat：具身人工智能研究平台", "date": "2019-04-02", "datePrecision": "day", "year": 2019, "url": "https://arxiv.org/abs/1904.01201", "paperUrl": "https://arxiv.org/abs/1904.01201", "codeUrl": "https://github.com/facebookresearch/habitat-lab", "summary": "Habitat把C++高速渲染/仿真后端与任务、传感器、回合和评估API分层，统一加载不同三维场景，使导航算法和数据集能独立替换。原论文通过PointGoal研究训练规模及跨场景迁移，主体不是新的具身基础模型，…", "abstractZh": "Habitat把C++高速渲染/仿真后端与任务、传感器、回合和评估API分层，统一加载不同三维场景，使导航算法和数据集能独立替换。原论文通过PointGoal研究训练规模及跨场景迁移，主体不是新的具身基础模型，也不涉及后来版本的完整操作物理系统。\n圆柱虚拟智能体以转10度、前进25厘米和停止动作导航，所有方法有理想GPS/罗盘；主实验执行无噪声，比较无视觉、RGB、Depth、RGBD的PPO与经典SLAM。Gibson/MP3D场景互斥划分，最多500步，距离目标0.2米内主动停止才成功；训练7500万步、五种子。", "category": "具身导航 / 仿真平台", "tags": [ "Habitat", "具身智能", "PointNav", "仿真基准" ], "directions": [ "数据集与基准", "导航与建图" ], "tier": "foundation", "experimentType": "sim", "robots": [], "robotFilters": [], "codeStatus": "open", "status": "实现已开源", "trainingNote": "官方 Habitat-Lab 含任务与强化学习基线，配合 Habitat-Sim 使用。代码与场景、任务数据、模型的许可证需要分别检查。", "contribution": "Habitat把C++高速渲染/仿真后端与任务、传感器、回合和评估API分层，统一加载不同三维场景，使导航算法和数据集能独立替换。原论文通过PointGoal研究训练规模及跨场景迁移，主体不是新的具身基础模型，也不涉及后来版本的完整操作物理系统。", "whyUseful": "复现需分别固定Habitat版本、场景授权、划分、动作碰撞行为、传感器和SPL定义，保留验证选模与五种子。不应将当前库接口或后续任务默认配置套回2019实验；应单列渲染基准硬件与学习网络吞吐。", "limitations": "任务有精确定位和深度，目标是坐标而非语言物体；没有实体机器人验证。RGB弱于Depth不证明颜色普遍无用；结果依赖任务、场景规模和训练预算。附录加入深度噪声后性能下降，渲染吞吐优势不等于感知/控制整体无瓶颈。", "license": "MIT（Habitat-Lab）；部分任务数据/模型受 CC BY-NC-SA 3.0 与场景条款约束", "evidence": [ { "url": "https://arxiv.org/abs/1904.01201", "note": "首版日期、Sim/API 架构、目标点导航与跨数据集实验。" }, { "url": "https://github.com/facebookresearch/habitat-lab", "note": "官方仓库确认训练基线、MIT 代码以及 Matterport3D/Gibson 数据和模型的独立条款。" }, { "url": "https://openaccess.thecvf.com/content_ICCV_2019/papers/Savva_Habitat_A_Platform_for_Embodied_AI_Research_ICCV_2019_paper.pdf", "note": "CVF 正式论文支持 2019 平台实验范围。" }, { "url": "https://arxiv.org/pdf/1904.01201v2", "note": "3 表1：Titan Xp+Xeon benchmark;128RGB five processes10592FPS;not end-to-end policy-training rate." }, { "url": "https://arxiv.org/pdf/1904.01201v2", "note": "4：Ideal GPS/Compass for every agent;noise-free actuation;0.2m success;500 actions." }, { "url": "https://arxiv.org/pdf/1904.01201v2", "note": "5 表2：75M steps,5 seeds;DepthSPL0.79/0.54;SLAM0.51/0.39." }, { "url": "https://arxiv.org/pdf/1904.01201v2", "note": "附录C：Noise-free-trained depth policy loses performance under synthetic depth noise;later scale test800M remains simulator-only." } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "以高吞吐仿真与统一任务接口，为具身导航的大规模学习和公平比较提供基础设施。", "experimentNote": "圆柱虚拟智能体以转10度、前进25厘米和停止动作导航，所有方法有理想GPS/罗盘；主实验执行无噪声，比较无视觉、RGB、Depth、RGBD的PPO与经典SLAM。Gibson/MP3D场景互斥划分，最多500步，距离目标0.2米内主动停止才成功；训练7500万步、五种子。", "projectUrl": null, "shortTitle": "", "original": { "id": "arxiv-1904.01201", "originalSourceUrl": "https://arxiv.org/pdf/1904.01201v2", "versionedOriginalUrl": "https://arxiv.org/pdf/1904.01201v2", "finalDownloadUrl": "https://arxiv.org/pdf/1904.01201", "sha256": "53485f2ed9af000b4f929be140c94e9bc097a59b69a0e5ae847446c0bb947e18", "bytes": 6108339, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:40:27.267804+00:00", "publicationDate": "2019/04/02", "publicationDateSourceUrl": "https://arxiv.org/abs/1904.01201", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1904.01201", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1904.01201", "pages": 17, "sourceIdentityStatus": "title_match", "sourceTitle": "Habitat: A Platform for Embodied AI Research", "sourceVersion": "1904.01201v2", "sourceVersionDate": "2019/11/25", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.406968+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1904.01201" }, "originalAnalysis": { "id": "arxiv-1904.01201", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739409+00:00", "sourceUrl": "https://arxiv.org/pdf/1904.01201v2", "sectionsRead": [ "3 Habitat-Sim/API架构及表1", "4 PointGoal任务、传感器与数据划分", "5 表2/图3–5 规模和跨数据集结果", "附录C 噪声深度与更长训练" ], "methodsZh": "Habitat把C++高速渲染/仿真后端与任务、传感器、回合和评估API分层，统一加载不同三维场景，使导航算法和数据集能独立替换。原论文通过PointGoal研究训练规模及跨场景迁移，主体不是新的具身基础模型，也不涉及后来版本的完整操作物理系统。", "experimentsZh": "圆柱虚拟智能体以转10度、前进25厘米和停止动作导航，所有方法有理想GPS/罗盘；主实验执行无噪声，比较无视觉、RGB、Depth、RGBD的PPO与经典SLAM。Gibson/MP3D场景互斥划分，最多500步，距离目标0.2米内主动停止才成功；训练7500万步、五种子。", "resultsZh": "作者表2报告Depth PPO在Gibson成功率89%、SPL0.79，在MP3D为69%、0.54；经典RGBD SLAM的SPL为0.51/0.39。训练时间较短时结论可能相反，跨数据集结果也显示来源难度影响固定算力下的表现。单场景128像素RGB多进程吞吐约10592帧/秒不是完整训练速率。", "limitationsZh": "任务有精确定位和深度，目标是坐标而非语言物体；没有实体机器人验证。RGB弱于Depth不证明颜色普遍无用；结果依赖任务、场景规模和训练预算。附录加入深度噪声后性能下降，渲染吞吐优势不等于感知/控制整体无瓶颈。", "reproductionZh": "复现需分别固定Habitat版本、场景授权、划分、动作碰撞行为、传感器和SPL定义，保留验证选模与五种子。不应将当前库接口或后续任务默认配置套回2019实验；应单列渲染基准硬件与学习网络吞吐。", "evidenceNotes": [ { "section/page": "3 表1", "note": "Titan Xp+Xeon benchmark;128RGB five processes10592FPS;not end-to-end policy-training rate." }, { "section/page": "4", "note": "Ideal GPS/Compass for every agent;noise-free actuation;0.2m success;500 actions." }, { "section/page": "5 表2", "note": "75M steps,5 seeds;DepthSPL0.79/0.54;SLAM0.51/0.39." }, { "section/page": "附录C", "note": "Noise-free-trained depth policy loses performance under synthetic depth noise;later scale test800M remains simulator-only." } ], "experimentType": "sim", "robots": [], "corrections": [ "2019原论文为理想GPS/罗盘下的仿真PointGoal，不能回填后续真实机器人能力。" ], "sourceVersion": "1904.01201v2", "originalSha256": "53485f2ed9af000b4f929be140c94e9bc097a59b69a0e5ae847446c0bb947e18" }, "analysisVerifiedAt": "2026-10-04T13:47:58.739409+00:00" }, { "id": "arxiv-1901.08652", "title": "Learning agile and dynamic motor skills for legged robots", "titleZh": "为足式机器人学习敏捷动态运动技能", "shortTitle": "Agile Motor Skills", "date": "2019-01-24", "datePrecision": "day", "year": 2019, "url": "https://arxiv.org/abs/1901.08652", "codeUrl": "https://github.com/junja94/anymal_science_robotics_supplementary", "category": "足式运动", "tags": [ "Sim2Real", "执行器模型", "四足" ], "tier": "classic", "summary": "用刚体接触仿真加由实体执行器数据训练的神经执行器模型，近似串联弹性电机、低层软件延迟和扭矩映射；再在随机化仿真中以强化学习训练读取状态历史的MLP，输出关节位置目标。行走、高速和跌倒恢复分别使用不同奖励训练。", "abstractZh": "用刚体接触仿真加由实体执行器数据训练的神经执行器模型，近似串联弹性电机、低层软件延迟和扭矩映射；再在随机化仿真中以强化学习训练读取状态历史的MLP，输出关节位置目标。行走、高速和跌倒恢复分别使用不同奖励训练。\n在真实ANYmal比较已有模型控制器的速度跟踪、机械功率和关节扭矩，另测试最高速度及九种跌倒初态；策略在桌面仿真训练后转移，部署单CPU推理约25微秒。研究同时检验不同执行器建模的转移效果。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "ANYmal" ], "codeStatus": "unknown", "status": "作者公开部分补充源码与数据；许可及完整训练系统未核实。", "trainingNote": "策略在仿真学习；执行器模型利用真实测量数据。", "whyUseful": "复现需ANYmal刚体/碰撞CAD、扭矩记录、执行器历史窗口、随机化和每项奖励，并保留恢复约束调整。论文仅称ANYmal，不能补成后续B/C型号；公开补充材料与内部仿真器的可用范围需另外核验。", "contribution": "用刚体接触仿真加由实体执行器数据训练的神经执行器模型，近似串联弹性电机、低层软件延迟和扭矩映射；再在随机化仿真中以强化学习训练读取状态历史的MLP，输出关节位置目标。行走、高速和跌倒恢复分别使用不同奖励训练。", "limitations": "仿真训练不等于零实机数据，执行器网络来自物理测量。恢复最初试验后放宽关节速度约束才得到报告的100%，不能称完全无硬件调试；精确模型、奖励安全成本和硬件约束仍有工程负担。", "caveats": "原论文只写ANYmal，不据后续项目推断B/C版本；策略仿真训练不等于零实机数据。", "evidence": [ { "url": "https://arxiv.org/abs/1901.08652", "note": "首发日期、ANYmal及速度跟踪/跌倒恢复。" }, { "url": "https://arxiv.org/pdf/1901.08652", "note": "执行器建模与仿真到真机流程。" }, { "url": "https://github.com/junja94/anymal_science_robotics_supplementary", "note": "作者补充仓库明确基于内部模拟器。" }, { "url": "https://github.com/junja94/anymal_science_robotics_supplementary/tree/master/actuator_model", "note": "公开MATLAB数据处理脚本和执行器训练数据。" }, { "url": "https://arxiv.org/pdf/1901.08652", "note": "Fig.1 / Methods：执行器网络从物理系统自监督训练。" }, { "url": "https://arxiv.org/pdf/1901.08652", "note": "High-speed locomotion：真机1.5m/s；仿真1.58m/s。" }, { "url": "https://arxiv.org/pdf/1901.08652", "note": "Discussion：放宽关节速度约束后恢复成功率达100%，为第二天实机结果。" } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "运动控制" ], "robotFilters": [ "ANYmal" ], "verificationNote": "2026-10-04核对论文与补充源码；为避免把公开片段等同完整开源，代码状态保留unknown。", "original": { "id": "arxiv-1901.08652", "originalSourceUrl": "https://arxiv.org/pdf/1901.08652v1", "versionedOriginalUrl": "https://arxiv.org/pdf/1901.08652v1", "finalDownloadUrl": "https://arxiv.org/pdf/1901.08652", "sha256": "c69eda2f422866925c3567ead246be231928af51dcdfa97f849222118d75ae45", "bytes": 7975310, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:40:49.265622+00:00", "publicationDate": "2019/01/24", "publicationDateSourceUrl": "https://arxiv.org/abs/1901.08652", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1901.08652", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1901.08652", "pages": 20, "sourceIdentityStatus": "title_match", "sourceTitle": "Learning agile and dynamic motor skills for legged robots", "sourceVersion": "1901.08652v1", "sourceVersionDate": "2019/01/24", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.417780+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1901.08652" }, "originalAnalysis": { "id": "arxiv-1901.08652", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090149+00:00", "sourceUrl": "https://arxiv.org/pdf/1901.08652", "sectionsRead": [ "Method overview / Fig.1", "Command-conditioned locomotion", "High-speed locomotion", "Recovery from a fall", "Discussion", "Figs.2–4" ], "methodsZh": "用刚体接触仿真加由实体执行器数据训练的神经执行器模型，近似串联弹性电机、低层软件延迟和扭矩映射；再在随机化仿真中以强化学习训练读取状态历史的MLP，输出关节位置目标。行走、高速和跌倒恢复分别使用不同奖励训练。", "experimentsZh": "在真实ANYmal比较已有模型控制器的速度跟踪、机械功率和关节扭矩，另测试最高速度及九种跌倒初态；策略在桌面仿真训练后转移，部署单CPU推理约25微秒。研究同时检验不同执行器建模的转移效果。", "resultsZh": "作者报告真实高速1.5米/秒、仿真1.58米/秒；给定命令下平均机械功率78.1瓦，对照97.3瓦。九种恢复初态全部成功，图示恢复少于三秒，体现动态接触策略的价值。", "limitationsZh": "仿真训练不等于零实机数据，执行器网络来自物理测量。恢复最初试验后放宽关节速度约束才得到报告的100%，不能称完全无硬件调试；精确模型、奖励安全成本和硬件约束仍有工程负担。", "reproductionZh": "复现需ANYmal刚体/碰撞CAD、扭矩记录、执行器历史窗口、随机化和每项奖励，并保留恢复约束调整。论文仅称ANYmal，不能补成后续B/C型号；公开补充材料与内部仿真器的可用范围需另外核验。", "experimentType": "both", "robots": [ "ANYmal" ], "evidenceNotes": [ { "section/page": "Fig.1 / Methods", "note": "执行器网络从物理系统自监督训练。" }, { "section/page": "High-speed locomotion", "note": "真机1.5m/s；仿真1.58m/s。" }, { "section/page": "Discussion", "note": "放宽关节速度约束后恢复成功率达100%，为第二天实机结果。" } ], "corrections": { "trainingNote": "控制策略仿真训练；执行器建模使用真实数据，且恢复约束有实机后调整。" } }, "experimentNote": "在真实ANYmal比较已有模型控制器的速度跟踪、机械功率和关节扭矩，另测试最高速度及九种跌倒初态；策略在桌面仿真训练后转移，部署单CPU推理约25微秒。研究同时检验不同执行器建模的转移效果。", "analysisVerifiedAt": "2026-10-04T13:50:35.090149+00:00" }, { "id": "arxiv-2403.06341", "title": "RTAB-Map as an Open-Source Lidar and Visual SLAM Library for Large-Scale and Long-Term Online Operation", "titleZh": "RTAB-Map：面向大规模长期在线运行的开源激光与视觉 SLAM 库", "date": "2019", "datePrecision": "year", "year": 2019, "url": "https://arxiv.org/abs/2403.06341", "paperUrl": "https://arxiv.org/abs/2403.06341", "codeUrl": "https://github.com/introlab/rtabmap", "summary": "RTAB-Map把多种视觉/激光里程计接入统一图SLAM，利用视觉词袋回环、邻近激光匹配和图优化，输出占据栅格或点云。短期、工作和长期记忆按计算/容量预算调度节点，重访时再载入旧位置，使在线计算不随完整地图无限…", "abstractZh": "RTAB-Map把多种视觉/激光里程计接入统一图SLAM，利用视觉词袋回环、邻近激光匹配和图优化，输出占据栅格或点云。短期、工作和长期记忆按计算/容量预算调度节点，重访时再载入旧位置，使在线计算不随完整地图无限增长。\n以0.16.3版本在KITTI、TUM RGB-D、EuRoC及MIT Stata Center真实记录序列比较传感器、里程计、轨迹误差与地图成本。长期实验连续播放两个PR2数据会话，2Hz地图更新，工作记忆限300节点，比较开关记忆管理。", "category": "多传感器 SLAM / 长期建图", "tags": [ "RTAB-Map", "内存管理", "RGB-D", "LiDAR", "长期运行", "真实数据评测" ], "directions": [ "导航与建图" ], "tier": "classic", "experimentType": "data", "robots": [ "PR2（MIT Stata Center数据采集平台）" ], "robotFilters": [ "PR2" ], "codeStatus": "open", "status": "实现已开源", "trainingNote": "经典估计与回环系统，无必需神经策略训练；官方 C++ 库、独立应用和 ROS 集成公开。依赖组件须分别核对授权。", "contribution": "RTAB-Map把多种视觉/激光里程计接入统一图SLAM，利用视觉词袋回环、邻近激光匹配和图优化，输出占据栅格或点云。短期、工作和长期记忆按计算/容量预算调度节点，重访时再载入旧位置，使在线计算不随完整地图无限增长。", "whyUseful": "应明确版本、传感输入、外部里程计和tf/时间同步，分别测在线最大漂移、最终误差和地图更新成本。不要把2024 arXiv上传日期替换原2019期刊年份；本次未运行ROS库。", "limitations": "记忆管理开启时在线全局地图并不含全部已访问环境，全部链接可事后重建；准确率对比此前关闭此机制。动态物体残影、标定、同步和几何退化仍影响导航。PR2是数据采集平台，不等于本文新闭环部署。", "license": "BSD-3-Clause（核心）；可选依赖许可证需另核", "evidence": [ { "url": "https://introlab.github.io/rtabmap/", "note": "作者项目页明确该论文发表于 JFR 36(2), 2019；arXiv 为 2024 后补上传，不作为原创年份。" }, { "url": "https://arxiv.org/abs/2403.06341", "note": "作者摘要核实 KITTI、EuRoC、TUM RGB-D、MIT Stata Center PR2 数据及多配置评估。" }, { "url": "https://github.com/introlab/rtabmap/blob/master/LICENSE", "note": "源码许可证三条 BSD 条款。" }, { "url": "https://arxiv.org/pdf/2403.06341", "note": "§3 Figure 1; §3.2：记忆结构与多传感同步。" }, { "url": "https://arxiv.org/pdf/2403.06341", "note": "§5.1 Figure 18; Tables 8–10：两会话回放、300节点、52毫秒开销及12厘米ATE。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "将长期地图内存管理与视觉/激光多配置评估结合，连接算法研究与机器人集成。", "experimentNote": "以0.16.3版本在KITTI、TUM RGB-D、EuRoC及MIT Stata Center真实记录序列比较传感器、里程计、轨迹误差与地图成本。长期实验连续播放两个PR2数据会话，2Hz地图更新，工作记忆限300节点，比较开关记忆管理。", "robotNote": "PR2 为论文所用真实数据集的机器人；其他平台未逐一核实。", "dateNote": "采用作者项目页引用的 2019 期刊年份；arXiv 在 2024-03-10 上传此文。", "projectUrl": null, "shortTitle": "", "original": { "id": "arxiv-2403.06341", "originalSourceUrl": "https://arxiv.org/pdf/2403.06341v1", "versionedOriginalUrl": "https://arxiv.org/pdf/2403.06341v1", "finalDownloadUrl": "https://arxiv.org/pdf/2403.06341", "sha256": "b6f641d5d2ed41e301940d57f1fb59a3cb6029f5ab04a21dff0c28013b19abe6", "bytes": 23360699, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:40:54.282655+00:00", "publicationDate": "2024/03/10", "publicationDateSourceUrl": "https://arxiv.org/abs/2403.06341", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/2403.06341", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/2403.06341", "pages": 40, "sourceIdentityStatus": "title_match", "sourceTitle": "RTAB-Map as an Open-Source Lidar and Visual SLAM Library for Large-Scale and Long-Term Online Operation", "sourceVersion": "2403.06341v1", "sourceVersionDate": "2024/03/10", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.436573+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/2403.06341" }, "originalAnalysis": { "id": "arxiv-2403.06341", "sourceUrl": "https://arxiv.org/pdf/2403.06341", "sectionsRead": [ "§3.1–3.6", "§4.1–4.4", "Tables 8–10", "§5.1", "§6–7" ], "methodsZh": "RTAB-Map把多种视觉/激光里程计接入统一图SLAM，利用视觉词袋回环、邻近激光匹配和图优化，输出占据栅格或点云。短期、工作和长期记忆按计算/容量预算调度节点，重访时再载入旧位置，使在线计算不随完整地图无限增长。", "experimentsZh": "以0.16.3版本在KITTI、TUM RGB-D、EuRoC及MIT Stata Center真实记录序列比较传感器、里程计、轨迹误差与地图成本。长期实验连续播放两个PR2数据会话，2Hz地图更新，工作记忆限300节点，比较开关记忆管理。", "resultsZh": "作者报告受限工作记忆使全部更新满足500毫秒预算，节点转移平均增加52毫秒，但其他模块计算减少；两种配置最终ATE均12厘米。MIT长走廊中短距纯激光易退化，融合轮速IMU显著改善，说明最佳传感组合依场景。", "limitationsZh": "记忆管理开启时在线全局地图并不含全部已访问环境，全部链接可事后重建；准确率对比此前关闭此机制。动态物体残影、标定、同步和几何退化仍影响导航。PR2是数据采集平台，不等于本文新闭环部署。", "reproductionZh": "应明确版本、传感输入、外部里程计和tf/时间同步，分别测在线最大漂移、最终误差和地图更新成本。不要把2024 arXiv上传日期替换原2019期刊年份；本次未运行ROS库。", "experimentType": "data", "robots": [ "PR2（MIT Stata Center数据采集平台）" ], "evidenceNotes": [ { "section/page": "§3 Figure 1; §3.2", "note": "记忆结构与多传感同步。" }, { "section/page": "§5.1 Figure 18; Tables 8–10", "note": "两会话回放、300节点、52毫秒开销及12厘米ATE。" } ], "corrections": [ { "field": "experimentType", "value": "data", "reason": "核心量化为真实传感器记录回放；PR2型号不代表本文闭环任务。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04" }, "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-1811.04551", "title": "Learning Latent Dynamics for Planning from Pixels", "titleZh": "从像素学习潜在动力学以进行规划", "shortTitle": "PlaNet", "date": "2018-11-12", "year": 2018, "url": "https://arxiv.org/abs/1811.04551", "paperUrl": "https://arxiv.org/abs/1811.04551", "projectUrl": "https://planetrl.github.io/", "codeUrl": "https://github.com/google-research/planet", "summary": "PlaNet从64×64图像学循环状态空间模型RSSM，用确定性记忆保存长期历史、随机潜变量表达未观测因素，同时重建图像和预测奖励。每步在潜空间用交叉熵法CEM反复筛选动作序列，只执行第一项后重规划，并持续收集…", "abstractZh": "PlaNet从64×64图像学循环状态空间模型RSSM，用确定性记忆保存长期历史、随机潜变量表达未观测因素，同时重建图像和预测奖励。每步在潜空间用交叉熵法CEM反复筛选动作序列，只执行第一项后重规划，并持续收集新经验更新模型。\nDeepMind Control六仿真任务含Cartpole、Reacher、Cheetah、Finger、Cup和Walker，涉及接触、遮挡与稀疏奖励；比较A3C/D4PG，消融纯确定/随机模型、随机采集、单次随机射击与多步潜变量目标，通常五种子。", "category": "世界模型 / 模型预测控制", "tags": [ "PlaNet", "世界模型", "强化学习" ], "directions": [ "世界模型", "强化学习" ], "tier": "foundation", "experimentType": "sim", "experimentNote": "DeepMind Control六仿真任务含Cartpole、Reacher、Cheetah、Finger、Cup和Walker，涉及接触、遮挡与稀疏奖励；比较A3C/D4PG，消融纯确定/随机模型、随机采集、单次随机射击与多步潜变量目标，通常五种子。", "robots": [], "robotFilters": [], "robotNote": "无实机；仿真形态不等同于已验证的商业机器人型号。", "codeStatus": "open", "status": "训练代码已开源", "trainingStatus": "训练代码已开源", "trainingNote": "官方训练代码为 Apache-2.0；使用较旧 TensorFlow 1.x 与 DMControl 依赖，复现需固定环境。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "Apache-2.0（代码）", "contribution": "PlaNet从64×64图像学循环状态空间模型RSSM，用确定性记忆保存长期历史、随机潜变量表达未观测因素，同时重建图像和预测奖励。每步在潜空间用交叉熵法CEM反复筛选动作序列，只执行第一项后重规划，并持续收集新经验更新模型。", "whyUseful": "先重现RSSM与CEM默认版，分别测超射目标和模型结构，不把两者混为一项改进；固定动作重复、预测窗口、采样候选与奖励处理，同时记录环境步数、墙钟训练时间和规划延迟。", "limitations": "无实机验证，样本效率比较依赖所选六任务与回合定义；模型误差和有限规划窗口限制控制。随机潜状态不是完整校准的不确定性保证，仍需环境奖励与持续在线交互。", "caveats": "需要任务奖励和在线交互；有限规划窗口及模型误差制约行为，原仓库已归档。", "evidence": [ { "url": "https://arxiv.org/abs/1811.04551", "note": "首发日期和核心方法。" }, { "url": "https://planetrl.github.io/", "note": "作者完整方法说明和视觉控制任务。" }, { "url": "https://github.com/google-research/planet", "note": "官方实现、训练命令、Apache-2.0 和归档状态。" }, { "url": "https://arxiv.org/pdf/1811.04551", "note": "2–5：RSSM+CEM、六任务及数据效率比较。" }, { "url": "https://arxiv.org/pdf/1811.04551", "note": "Appendix D Fig.8：超射目标对DRNN与RSSM的相反影响。" }, { "url": "https://arxiv.org/pdf/1811.04551v5", "note": "2–5：RSSM+CEM、六任务及数据效率比较。" }, { "url": "https://arxiv.org/pdf/1811.04551v5", "note": "Appendix D Fig.8：超射目标对DRNN与RSSM的相反影响。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2018：RSSM 与潜空间 MPC 为后续 Dreamer 系列奠定结构基础。", "freshness": "基础奠基", "original": { "id": "arxiv-1811.04551", "originalSourceUrl": "https://arxiv.org/pdf/1811.04551v5", "versionedOriginalUrl": "https://arxiv.org/pdf/1811.04551v5", "finalDownloadUrl": "https://arxiv.org/pdf/1811.04551", "sha256": "abac727526e6a45669d3ab9957126587e22aacf4dc9bcd60ffe5c853108e5bcc", "bytes": 3146899, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:40:53.602154+00:00", "publicationDate": "2018/11/12", "publicationDateSourceUrl": "https://arxiv.org/abs/1811.04551", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1811.04551", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1811.04551", "pages": 20, "sourceIdentityStatus": "title_match", "sourceTitle": "Learning Latent Dynamics for Planning from Pixels", "sourceVersion": "1811.04551v5", "sourceVersionDate": "2019/06/04", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.440825+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1811.04551" }, "originalAnalysis": { "id": "arxiv-1811.04551", "sourceUrl": "https://arxiv.org/pdf/1811.04551v5", "sectionsRead": [ "2", "3", "4", "5", "7", "Appendix A,D" ], "methodsZh": "PlaNet从64×64图像学循环状态空间模型RSSM，用确定性记忆保存长期历史、随机潜变量表达未观测因素，同时重建图像和预测奖励。每步在潜空间用交叉熵法CEM反复筛选动作序列，只执行第一项后重规划，并持续收集新经验更新模型。", "experimentsZh": "DeepMind Control六仿真任务含Cartpole、Reacher、Cheetah、Finger、Cup和Walker，涉及接触、遮挡与稀疏奖励；比较A3C/D4PG，消融纯确定/随机模型、随机采集、单次随机射击与多步潜变量目标，通常五种子。", "resultsZh": "作者报告接近强无模型基线的成绩而平均少约200倍回合，单V100训练约10–20小时；在线有目标采集及CEM迭代均重要。附录D明确：latent overshooting帮助DRNN，却略降低RSSM表现，不能把它概括为PlaNet必需的增益。", "limitationsZh": "无实机验证，样本效率比较依赖所选六任务与回合定义；模型误差和有限规划窗口限制控制。随机潜状态不是完整校准的不确定性保证，仍需环境奖励与持续在线交互。", "reproductionZh": "先重现RSSM与CEM默认版，分别测超射目标和模型结构，不把两者混为一项改进；固定动作重复、预测窗口、采样候选与奖励处理，同时记录环境步数、墙钟训练时间和规划延迟。", "experimentType": "sim", "robots": [], "corrections": [ "附录D中latent overshooting略降低RSSM成绩；收益主要出现在DRNN等模型，不能一概声称提升主模型。" ], "evidenceNotes": [ { "note": "RSSM+CEM、六任务及数据效率比较。", "section": "2–5" }, { "note": "超射目标对DRNN与RSSM的相反影响。", "section": "Appendix D Fig.8" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "1811.04551v5", "originalSha256": "abac727526e6a45669d3ab9957126587e22aacf4dc9bcd60ffe5c853108e5bcc" }, "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-1808.00177", "title": "Learning Dexterous In-Hand Manipulation", "titleZh": "学习灵巧的手内操作", "shortTitle": "Dactyl / Learning Dexterity", "date": "2018-08-01", "datePrecision": "day", "year": 2018, "url": "https://arxiv.org/abs/1808.00177", "paperUrl": "https://arxiv.org/abs/1808.00177", "codeUrl": null, "summary": "Dactyl用循环PPO和非对称actor-critic学习Shadow手内物体重定向；训练时随机化动力学、观察与动作噪声及未建模效应，借记忆适应每回合不同参数。控制策略接收指尖位置和物体目标误差，输出分成11…", "abstractZh": "Dactyl用循环PPO和非对称actor-critic学习Shadow手内物体重定向；训练时随机化动力学、观察与动作噪声及未建模效应，借记忆适应每回合不同参数。控制策略接收指尖位置和物体目标误差，输出分成11档的相对关节目标；视觉网络另用随机化合成图估计物体位置姿态。\n任务对象为方块及八棱柱，每成功一次就换目标，掉落、超时或累计50次结束。每策略仿真100次、实机10次，并比较锁腕、视觉/动捕物体估计、随机化组及是否有记忆。Shadow手24自由度、20组驱动，12Hz策略经约1kHz低层执行。", "category": "灵巧手 / Sim2Real", "tags": [ "手内重定向", "域随机化", "视觉控制", "强化学习" ], "directions": [ "灵巧手", "操作与抓取", "强化学习" ], "tier": "classic", "experimentType": "both", "robots": [ "Shadow Dexterous Hand" ], "robotNote": "", "codeStatus": "unknown", "status": "本次未核实2018原实验的完整官方训练代码。", "trainingNote": "分布式RL在随机化仿真中训练控制策略，视觉估计器与硬件系统另行构建；后续robogym不能直接视为本篇完整复现包。", "contribution": "Dactyl用循环PPO和非对称actor-critic学习Shadow手内物体重定向；训练时随机化动力学、观察与动作噪声及未建模效应，借记忆适应每回合不同参数。控制策略接收指尖位置和物体目标误差，输出分成11档的相对关节目标；视觉网络另用随机化合成图估计物体位置姿态。", "whyUseful": "默认训练资源为8GPU优化器加6144个CPU采样核，完全随机化达到相似仿真表现约需百年模拟经验、50小时墙钟。复现需手部校准、追踪、三相机、Unity视觉渲染及低层控制，不可把后续robogym当作完整原实验包。", "limitations": "视觉版仍用PhaseSpace追踪五个指尖，且实验采用白背景并清洁物体；不使用触觉输入。对象有限，十次试验方差大，50次是截断上限；作者还报告硬件损坏影响实验。手从预放物体开始，不是任意抓取与操作。", "evidence": [ { "url": "https://arxiv.org/abs/1808.00177", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/pdf/1808.00177", "note": "原文核对Shadow手、仿真训练与真实实验范围。" }, { "url": "https://openai.com/index/learning-dexterity/", "note": "研究团队的项目说明与实验展示。" }, { "url": "https://arxiv.org/pdf/1808.00177v5", "note": "2.1/3.1：Shadow24DoF/20actuators;PhaseSpace fingertip tracking even in vision object-pose setup;no tactile input." }, { "url": "https://arxiv.org/pdf/1808.00177v5", "note": "6.2 表3：100sim/10real per policy;cap50;state18.8mean/13median;vision15.2mean/11.5median." }, { "url": "https://arxiv.org/pdf/1808.00177v5", "note": "6.3 表4：Without all randomization real median0;without physics median2." }, { "url": "https://arxiv.org/pdf/1808.00177v5", "note": "6.3/6.5：~100years simulated experience≈50h;default8GPU+6144CPU rollout cores." }, { "url": "https://arxiv.org/pdf/1808.00177v5", "note": "v5 PDF p10/Table3 visual check：Confirmed10real/100sim trials,80-second per-goal timeout,50-rotation cap;authors note white background/object cleaning for vision and hardware breakages." } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2018：大规模仿真RL跨越到真实灵巧手，推动域随机化的机器人应用。", "projectUrl": null, "robotFilters": [ "Shadow Hand" ], "original": { "id": "arxiv-1808.00177", "originalSourceUrl": "https://arxiv.org/pdf/1808.00177v5", "versionedOriginalUrl": "https://arxiv.org/pdf/1808.00177v5", "finalDownloadUrl": "https://arxiv.org/pdf/1808.00177", "sha256": "da6ec44dd0941d674ec58334ff630f83d7bd19e6062ab2837ccfa0b3b94f6f73", "bytes": 4590851, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:01.115298+00:00", "publicationDate": "2018/08/01", "publicationDateSourceUrl": "https://arxiv.org/abs/1808.00177", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1808.00177", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1808.00177", "pages": 27, "sourceIdentityStatus": "title_match", "sourceTitle": "Learning Dexterous In-Hand Manipulation", "sourceVersion": "1808.00177v5", "sourceVersionDate": "2019/01/18", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.446696+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1808.00177" }, "originalAnalysis": { "id": "arxiv-1808.00177", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739410+00:00", "sourceUrl": "https://arxiv.org/pdf/1808.00177v5", "sectionsRead": [ "2 硬件、MuJoCo/Unity与任务", "3 随机化和观察", "4–5 LSTM/PPO与视觉估计", "6.2–6.6 表3–6、消融、规模", "归档v5 PDF第10页表3可视复核" ], "methodsZh": "Dactyl用循环PPO和非对称actor-critic学习Shadow手内物体重定向；训练时随机化动力学、观察与动作噪声及未建模效应，借记忆适应每回合不同参数。控制策略接收指尖位置和物体目标误差，输出分成11档的相对关节目标；视觉网络另用随机化合成图估计物体位置姿态。", "experimentsZh": "任务对象为方块及八棱柱，每成功一次就换目标，掉落、超时或累计50次结束。每策略仿真100次、实机10次，并比较锁腕、视觉/动捕物体估计、随机化组及是否有记忆。Shadow手24自由度、20组驱动，12Hz策略经约1kHz低层执行。", "resultsZh": "作者表3报告实机方块状态输入连续成功均值18.8、中位13，视觉输入均值15.2、中位11.5；仿真状态均值43.4，现实差距仍明显。去全部随机化后实机中位数为0，去物理随机化为2，支持强随机化而非单靠理想仿真。", "limitationsZh": "视觉版仍用PhaseSpace追踪五个指尖，且实验采用白背景并清洁物体；不使用触觉输入。对象有限，十次试验方差大，50次是截断上限；作者还报告硬件损坏影响实验。手从预放物体开始，不是任意抓取与操作。", "reproductionZh": "默认训练资源为8GPU优化器加6144个CPU采样核，完全随机化达到相似仿真表现约需百年模拟经验、50小时墙钟。复现需手部校准、追踪、三相机、Unity视觉渲染及低层控制，不可把后续robogym当作完整原实验包。", "evidenceNotes": [ { "section/page": "2.1/3.1", "note": "Shadow24DoF/20actuators;PhaseSpace fingertip tracking even in vision object-pose setup;no tactile input." }, { "section/page": "6.2 表3", "note": "100sim/10real per policy;cap50;state18.8mean/13median;vision15.2mean/11.5median." }, { "section/page": "6.3 表4", "note": "Without all randomization real median0;without physics median2." }, { "section/page": "6.3/6.5", "note": "~100years simulated experience≈50h;default8GPU+6144CPU rollout cores." }, { "section/page": "v5 PDF p10/Table3 visual check", "note": "Confirmed10real/100sim trials,80-second per-goal timeout,50-rotation cap;authors note white background/object cleaning for vision and hardware breakages." } ], "experimentType": "both", "robots": [ "Shadow Dexterous Hand" ], "corrections": [ "视觉物体估计版本仍依赖指尖动捕，不能称全视觉无外部追踪。" ], "sourceVersion": "1808.00177v5", "originalSha256": "da6ec44dd0941d674ec58334ff630f83d7bd19e6062ab2837ccfa0b3b94f6f73" }, "experimentNote": "任务对象为方块及八棱柱，每成功一次就换目标，掉落、超时或累计50次结束。每策略仿真100次、实机10次，并比较锁腕、视觉/动捕物体估计、随机化组及是否有记忆。Shadow手24自由度、20组驱动，12Hz策略经约1kHz低层执行。", "analysisVerifiedAt": "2026-10-04T13:47:58.739410+00:00" }, { "id": "arxiv-1806.10293", "title": "QT-Opt: Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation", "titleZh": "QT-Opt：面向视觉机器人操作的可扩展深度强化学习", "shortTitle": "QT-Opt", "date": "2018-06-27", "datePrecision": "day", "year": 2018, "url": "https://arxiv.org/abs/1806.10293", "codeUrl": "https://github.com/google-research/tensor2robot/tree/master/research/qtopt", "category": "强化学习", "tags": [ "抓取", "离策略强化学习", "大规模数据" ], "tier": "classic", "summary": "QT-Opt以连续动作Q学习替代独立actor，通过交叉熵方法随机优化Q值选择动作；双目标网络、裁剪双Q和异步Bellman更新支撑大规模离策略图像数据。闭环反复观察，使抓取前推挪、重抓和失败恢复可由长期奖励学…", "abstractZh": "QT-Opt以连续动作Q学习替代独立actor，通过交叉熵方法随机优化Q值选择动作；双目标网络、裁剪双Q和异步Bellman更新支撑大规模离策略图像数据。闭环反复观察，使抓取前推挪、重抓和失败恢复可由长期奖励学习。\n七台KUKA LBR IIWA从多轮实验累积58万离策略抓取，后加2.8万在线抓取联合微调。未见物体测试每机器人102次、成功物体放回；清空箱测试单机器人28物体、最多30次尝试，重复五轮，附录另有仿真消融。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "KUKA LBR IIWA" ], "codeStatus": "open", "status": "官方部分开源：critic网络；Bellman目标流程与真实数据未公开。", "trainingNote": "真实数据大规模离策略训练，并结合在线数据微调。", "whyUseful": "复现须保留放回/不放回协议、未见物体集合、脚本和在线数据比例、分布式目标更新及动作空间。公开critic示例不能代替原始真实数据与完整Bellman系统；本轮未训练或执行抓取。", "contribution": "QT-Opt以连续动作Q学习替代独立actor，通过交叉熵方法随机优化Q值选择动作；双目标网络、裁剪双Q和异步Bellman更新支撑大规模离策略图像数据。闭环反复观察，使抓取前推挪、重抓和失败恢复可由长期奖励学习。", "limitations": "数据来自多次实验复用，作者承认真实系统难做严格受控训练对照；初期探索由15–30%成功的弱脚本启动，非纯随机。任务限于箱内抓取，图中腕相机没有用于实验，且大规模机器人数据成本高。", "caveats": "论文的96%是指定未见物体评测结果，不能理解为任意物体或场景的保证。", "evidence": [ { "url": "https://arxiv.org/abs/1806.10293", "note": "首发日期、数据规模与摘要。" }, { "url": "https://arxiv.org/pdf/1806.10293", "note": "图2明确KUKA LBR IIWA；正文真实实验和附录仿真。" }, { "url": "https://github.com/google-research/tensor2robot/tree/master/research/qtopt", "note": "README明确仅网络架构、mock data示例；Bellman目标进程未开源。" }, { "url": "https://arxiv.org/pdf/1806.10293", "note": "§5：7台、四个月约800机器人小时；复用实验数据。" }, { "url": "https://arxiv.org/pdf/1806.10293", "note": "§6.1 / Table 1：96%放回；清箱30次76%，2/5完全清空。" }, { "url": "https://arxiv.org/pdf/1806.10293", "note": "§6.1 footnote：腕部相机未使用，视觉来自肩后RGB。" }, { "url": "https://arxiv.org/pdf/1806.10293", "note": "PDF p.7, Table 1 (visual inspection)：已核对96%位于Test放回列，清箱30次为76%，两种协议不可互换。" } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "操作与抓取", "强化学习" ], "robotFilters": [ "KUKA LBR IIWA" ], "verificationNote": "2026-10-04核对论文与官方代码范围；未将部分开源等同完整可复现。", "original": { "id": "arxiv-1806.10293", "originalSourceUrl": "https://arxiv.org/pdf/1806.10293v3", "versionedOriginalUrl": "https://arxiv.org/pdf/1806.10293v3", "finalDownloadUrl": "https://arxiv.org/pdf/1806.10293", "sha256": "1482043a0e94f30c0db5b1841574a81a44eeaa47154ca99492e31c3706b89c75", "bytes": 5491270, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:05.494749+00:00", "publicationDate": "2018/06/27", "publicationDateSourceUrl": "https://arxiv.org/abs/1806.10293", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1806.10293", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1806.10293", "pages": 23, "sourceIdentityStatus": "title_match", "sourceTitle": "QT-Opt: Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation", "sourceVersion": "1806.10293v3", "sourceVersionDate": "2018/11/28", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.452254+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1806.10293" }, "originalAnalysis": { "id": "arxiv-1806.10293", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090150+00:00", "sourceUrl": "https://arxiv.org/pdf/1806.10293", "sectionsRead": [ "§4–5", "§6.1–6.2", "Table 1", "Appendices A,C,F" ], "methodsZh": "QT-Opt以连续动作Q学习替代独立actor，通过交叉熵方法随机优化Q值选择动作；双目标网络、裁剪双Q和异步Bellman更新支撑大规模离策略图像数据。闭环反复观察，使抓取前推挪、重抓和失败恢复可由长期奖励学习。", "experimentsZh": "七台KUKA LBR IIWA从多轮实验累积58万离策略抓取，后加2.8万在线抓取联合微调。未见物体测试每机器人102次、成功物体放回；清空箱测试单机器人28物体、最多30次尝试，重复五轮，附录另有仿真消融。", "resultsZh": "作者报告放回测试成功96%，只离策略训练87%，既有系统78%。不放回清箱前30次成功76%，五轮中两轮在30次内清空；因此96%不能用于描述所有箱子清空或最后难物体的可靠性。", "limitationsZh": "数据来自多次实验复用，作者承认真实系统难做严格受控训练对照；初期探索由15–30%成功的弱脚本启动，非纯随机。任务限于箱内抓取，图中腕相机没有用于实验，且大规模机器人数据成本高。", "reproductionZh": "复现须保留放回/不放回协议、未见物体集合、脚本和在线数据比例、分布式目标更新及动作空间。公开critic示例不能代替原始真实数据与完整Bellman系统；本轮未训练或执行抓取。", "experimentType": "both", "robots": [ "KUKA LBR IIWA" ], "evidenceNotes": [ { "section/page": "§5", "note": "7台、四个月约800机器人小时；复用实验数据。" }, { "section/page": "§6.1 / Table 1", "note": "96%放回；清箱30次76%，2/5完全清空。" }, { "section/page": "§6.1 footnote", "note": "腕部相机未使用，视觉来自肩后RGB。" }, { "section/page": "PDF p.7, Table 1 (visual inspection)", "note": "已核对96%位于Test放回列，清箱30次为76%，两种协议不可互换。" } ] }, "experimentNote": "七台KUKA LBR IIWA从多轮实验累积58万离策略抓取，后加2.8万在线抓取联合微调。未见物体测试每机器人102次、成功物体放回；清空箱测试单机器人28物体、最多30次尝试，重复五轮，附录另有仿真消融。", "analysisVerifiedAt": "2026-10-04T13:50:35.090150+00:00" }, { "id": "arxiv-1804.02717", "title": "DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills", "titleZh": "DeepMimic：由动作示范引导的物理角色技能深度强化学习", "shortTitle": "DeepMimic", "date": "2018-04-08", "datePrecision": "day", "year": 2018, "url": "https://arxiv.org/abs/1804.02717", "paperUrl": "https://arxiv.org/abs/1804.02717", "codeUrl": "https://github.com/xbpeng/DeepMimic", "summary": "以动作示范的姿态、关节速度、末端和质心相似奖励，配合任务奖励训练PPO物理控制器。随机从参考动作状态初始化并在摔倒时提前终止，帮助探索高动态技能；策略输出PD目标，多片段奖励、选择器或价值门控支持技能组合。", "abstractZh": "以动作示范的姿态、关节速度、末端和质心相似奖励，配合任务奖励训练PPO物理控制器。随机从参考动作状态初始化并在摔倒时提前终止，帮助探索高动态技能；策略输出PD目标，多片段奖励、选择器或价值门控支持技能组合。\nBullet中测试人形、Atlas、霸王龙和龙等虚拟角色，30Hz策略、1.2kHz物理仿真。动作参考多为0.5–5秒捕捉片段；比较单纯模仿、单纯任务及两者结合，消融参考初始化/提前终止，并测推力扰动容忍。", "category": "运动控制 / 动作模仿", "tags": [ "动作捕捉", "运动跟踪", "强化学习", "物理动画" ], "directions": [ "运动控制", "基础理论", "强化学习", "模仿学习" ], "tier": "foundation", "experimentType": "sim", "robots": [ "Atlas（仿真）" ], "robotNote": "Atlas仅为仿真角色；另有人类、恐龙和龙等虚拟角色，不作为实机型号。", "codeStatus": "open", "status": "作者训练与演示代码公开；旧仓库已标注弃用。", "trainingNote": "使用物理仿真、动作片段和任务回报训练策略；原仓库依赖Bullet与旧版TensorFlow，作者建议新项目参考MimicKit。", "contribution": "以动作示范的姿态、关节速度、末端和质心相似奖励，配合任务奖励训练PPO物理控制器。随机从参考动作状态初始化并在摔倒时提前终止，帮助探索高动态技能；策略输出PD目标，多片段奖励、选择器或价值门控支持技能组合。", "whyUseful": "需对齐参考动作重定向、相位、初始化分布、接触终止和低层PD设置，既评任务成功也检查动作形态。把100次评测与训练种子数分开报告，本次未运行Bullet训练。", "limitations": "全部为物理动画，Atlas不代表实机。多数训练统计来自单次训练，技能常需数天；阶段变量随时间线性推进限制灵活时序，大动作库整合未验证。PD参数与奖励权重需人工选择，部署机器人仍属未来工作。", "evidence": [ { "url": "https://arxiv.org/abs/1804.02717", "note": "arXiv原始记录：标题、首发日期与方法摘要。" }, { "url": "https://xbpeng.github.io/projects/DeepMimic/index.html", "note": "作者项目页明确角色类型、仿真范围及官方代码入口。" }, { "url": "https://github.com/xbpeng/DeepMimic", "note": "README明确同时支持DeepMimic/AMP，给出训练流程和弃用提示。" }, { "url": "https://arxiv.org/pdf/1804.02717", "note": "§5–7; Figure 3：参考模仿、RSI/ET及虚拟角色。" }, { "url": "https://arxiv.org/pdf/1804.02717", "note": "Table 4; §10.4–11：踢击99/19/55%、单训练运行和真实部署未验证。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对原论文、作者项目页与列出的代码证据；未运行训练或独立复现实验。", "timelineNote": "2018：示范驱动的物理运动控制成为后续人形模仿学习的重要技术起点。", "projectUrl": null, "robotFilters": [], "original": { "id": "arxiv-1804.02717", "originalSourceUrl": "https://arxiv.org/pdf/1804.02717v3", "versionedOriginalUrl": "https://arxiv.org/pdf/1804.02717v3", "finalDownloadUrl": "https://arxiv.org/pdf/1804.02717", "sha256": "4bd4cc353bd2dc19da99bb21a254cbfdb9840e49940ce69d484eb6d45bfa348f", "bytes": 7770304, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:05.631393+00:00", "publicationDate": "2018/04/08", "publicationDateSourceUrl": "https://arxiv.org/abs/1804.02717", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1804.02717", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1804.02717", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills", "sourceVersion": "1804.02717v3", "sourceVersionDate": "2018/07/27", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.460323+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1804.02717" }, "originalAnalysis": { "id": "arxiv-1804.02717", "sourceUrl": "https://arxiv.org/pdf/1804.02717", "sectionsRead": [ "§5–7", "§10 Tables 3–6", "§11" ], "methodsZh": "以动作示范的姿态、关节速度、末端和质心相似奖励，配合任务奖励训练PPO物理控制器。随机从参考动作状态初始化并在摔倒时提前终止，帮助探索高动态技能；策略输出PD目标，多片段奖励、选择器或价值门控支持技能组合。", "experimentsZh": "Bullet中测试人形、Atlas、霸王龙和龙等虚拟角色，30Hz策略、1.2kHz物理仿真。动作参考多为0.5–5秒捕捉片段；比较单纯模仿、单纯任务及两者结合，消融参考初始化/提前终止，并测推力扰动容忍。", "resultsZh": "作者报告踢击任务联合目标成功99%，仅模仿19%、仅任务55%，各100次。参考初始化缺失时后空翻可能退化成小跳；归一化回报相似也未必意味着复现正确动作。学得步态可随目标调整并在仿真中抗扰动。", "limitationsZh": "全部为物理动画，Atlas不代表实机。多数训练统计来自单次训练，技能常需数天；阶段变量随时间线性推进限制灵活时序，大动作库整合未验证。PD参数与奖励权重需人工选择，部署机器人仍属未来工作。", "reproductionZh": "需对齐参考动作重定向、相位、初始化分布、接触终止和低层PD设置，既评任务成功也检查动作形态。把100次评测与训练种子数分开报告，本次未运行Bullet训练。", "experimentType": "sim", "robots": [ "Atlas（仿真）" ], "evidenceNotes": [ { "section/page": "§5–7; Figure 3", "note": "参考模仿、RSI/ET及虚拟角色。" }, { "section/page": "Table 4; §10.4–11", "note": "踢击99/19/55%、单训练运行和真实部署未验证。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "Bullet中测试人形、Atlas、霸王龙和龙等虚拟角色，30Hz策略、1.2kHz物理仿真。动作参考多为0.5–5秒捕捉片段；比较单纯模仿、单纯任务及两者结合，消融参考初始化/提前终止，并测推力扰动容忍。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-1803.10122", "title": "World Models", "titleZh": "世界模型", "shortTitle": "World Models", "date": "2018-03-27", "year": 2018, "url": "https://arxiv.org/abs/1803.10122", "paperUrl": "https://arxiv.org/abs/1803.10122", "projectUrl": "https://worldmodels.github.io/", "codeUrl": "https://github.com/hardmaru/WorldModelsExperiments", "summary": "World Models把感知VAE、预测潜状态分布的MDN-RNN和小线性控制器分开训练。先用随机策略采集一万条轨迹训练V/M，再由CMA-ES优化控制器；动作依据潜变量及循环记忆，部分实验把优化过程移到模型…", "abstractZh": "World Models把感知VAE、预测潜状态分布的MDN-RNN和小线性控制器分开训练。先用随机策略采集一万条轨迹训练V/M，再由CMA-ES优化控制器；动作依据潜变量及循环记忆，部分实验把优化过程移到模型生成的潜在环境。\nCarRacing-v0检验原游戏环境内训练的控制器，比较仅视觉和加入循环记忆；VizDoom Take Cover则在“梦境”中训练后迁回真实游戏引擎，改变MDN采样温度研究模型漏洞。两者都是游戏模拟，文中real environment不是现实机器人。", "category": "世界模型 / 潜在动力学", "tags": [ "World Models", "世界模型", "强化学习" ], "directions": [ "世界模型", "强化学习" ], "tier": "foundation", "experimentType": "sim", "experimentNote": "CarRacing-v0检验原游戏环境内训练的控制器，比较仅视觉和加入循环记忆；VizDoom Take Cover则在“梦境”中训练后迁回真实游戏引擎，改变MDN采样温度研究模型漏洞。两者都是游戏模拟，文中real environment不是现实机器人。", "robots": [], "robotFilters": [], "robotNote": "无实机；仿真形态不等同于已验证的商业机器人型号。", "codeStatus": "unknown", "status": "代码公开，许可未核实", "trainingStatus": "代码公开，许可未核实", "trainingNote": "作者提供 TensorFlow 参考实验及复现教程；仓库顶层许可未核实，不能据公开代码推断完整开源授权。", "codeStatusNote": "代码许可和训练范围以本条说明为准；公开仓库不代表全部数据、权重和依赖拥有同一许可。", "license": "未核实", "contribution": "World Models把感知VAE、预测潜状态分布的MDN-RNN和小线性控制器分开训练。先用随机策略采集一万条轨迹训练V/M，再由CMA-ES优化控制器；动作依据潜变量及循环记忆，部分实验把优化过程移到模型生成的潜在环境。", "whyUseful": "分别复现CarRacing与Doom训练场所、终止模型、采样温度及CMA-ES种群设置；报告原游戏验证而非仅想象奖励。固定历史依赖版本和随机种子，检查游戏步数与真实机器人时间不可互换。", "limitations": "控制器在CarRacing并非完全在梦境训练；模型由随机探索数据训练，未覆盖的新状态易产生幻觉。较高温度也会使学习过难，潜特征可能重建无关细节而遗漏任务信息；示例无法证明一般世界建模能力。", "caveats": "结果限于游戏环境；模型误差可能被策略利用；原始依赖较旧。", "evidence": [ { "url": "https://arxiv.org/abs/1803.10122", "note": "原论文与首发日期。" }, { "url": "https://worldmodels.github.io/", "note": "作者交互论文：VAE、MDN-RNN、控制器及两类游戏实验。" }, { "url": "https://github.com/hardmaru/WorldModelsExperiments", "note": "作者参考实现；许可未核实。" }, { "url": "https://arxiv.org/pdf/1803.10122", "note": "2–3 Table 1：V/M/C分工与CarRacing906±21。" }, { "url": "https://arxiv.org/pdf/1803.10122", "note": "4 Table 2; 7：Doom温度实验、模型漏洞及表征限制。" }, { "url": "https://arxiv.org/pdf/1803.10122v4", "note": "2–3 Table 1：V/M/C分工与CarRacing906±21。" }, { "url": "https://arxiv.org/pdf/1803.10122v4", "note": "4 Table 2; 7：Doom温度实验、模型漏洞及表征限制。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2018：以可复现实验普及“学习环境模型、在想象中训练控制器”的现代深度学习范式。", "freshness": "基础奠基", "original": { "id": "arxiv-1803.10122", "originalSourceUrl": "https://arxiv.org/pdf/1803.10122v4", "versionedOriginalUrl": "https://arxiv.org/pdf/1803.10122v4", "finalDownloadUrl": "https://arxiv.org/pdf/1803.10122", "sha256": "b0c1e30aab53efd28ddf61d661f680150918d4d03b77bae62bc52d62dbd76cce", "bytes": 3147467, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:12.795754+00:00", "publicationDate": "2018/03/27", "publicationDateSourceUrl": "https://arxiv.org/abs/1803.10122", "license": "http://creativecommons.org/licenses/by/4.0/", "licenseSourceUrl": "https://arxiv.org/abs/1803.10122", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1803.10122", "pages": 21, "sourceIdentityStatus": "title_match", "sourceTitle": "World Models", "sourceVersion": "1803.10122v4", "sourceVersionDate": "2018/05/09", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.466480+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1803.10122" }, "originalAnalysis": { "id": "arxiv-1803.10122", "sourceUrl": "https://arxiv.org/pdf/1803.10122v4", "sectionsRead": [ "2.1–2.3", "3.1–3.3", "4.1–4.3", "5", "7", "Tables 1–2" ], "methodsZh": "World Models把感知VAE、预测潜状态分布的MDN-RNN和小线性控制器分开训练。先用随机策略采集一万条轨迹训练V/M，再由CMA-ES优化控制器；动作依据潜变量及循环记忆，部分实验把优化过程移到模型生成的潜在环境。", "experimentsZh": "CarRacing-v0检验原游戏环境内训练的控制器，比较仅视觉和加入循环记忆；VizDoom Take Cover则在“梦境”中训练后迁回真实游戏引擎，改变MDN采样温度研究模型漏洞。两者都是游戏模拟，文中real environment不是现实机器人。", "resultsZh": "CarRacing100次随机赛道平均906±21，达到其900分解决标准；Doom温度1.15时原游戏生存1092±556步，而随机策略210±108。过低温度会让策略找到使火球消失等模型漏洞，适度随机性有助减少利用不实动力学。", "limitationsZh": "控制器在CarRacing并非完全在梦境训练；模型由随机探索数据训练，未覆盖的新状态易产生幻觉。较高温度也会使学习过难，潜特征可能重建无关细节而遗漏任务信息；示例无法证明一般世界建模能力。", "reproductionZh": "分别复现CarRacing与Doom训练场所、终止模型、采样温度及CMA-ES种群设置；报告原游戏验证而非仅想象奖励。固定历史依赖版本和随机种子，检查游戏步数与真实机器人时间不可互换。", "experimentType": "sim", "robots": [], "corrections": [ "CarRacing控制器在原环境优化；Doom才展示梦境内训练再回原游戏，两者都不是机器人sim-to-real。" ], "evidenceNotes": [ { "note": "V/M/C分工与CarRacing906±21。", "section": "2–3 Table 1" }, { "note": "Doom温度实验、模型漏洞及表征限制。", "section": "4 Table 2; 7" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "1803.10122v4", "originalSha256": "b0c1e30aab53efd28ddf61d661f680150918d4d03b77bae62bc52d62dbd76cce" }, "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-1801.01290", "title": "Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor", "titleZh": "Soft Actor-Critic：带随机策略的离策略最大熵深度强化学习", "shortTitle": "SAC", "date": "2018-01-04", "datePrecision": "day", "year": 2018, "url": "https://arxiv.org/abs/1801.01290", "codeUrl": "https://github.com/haarnoja/sac", "category": "强化学习", "tags": [ "最大熵", "离策略强化学习", "基础算法" ], "tier": "foundation", "summary": "原版SAC将奖励与策略熵共同优化，以离策略回放提高利用率。软策略迭代先评估软Q值，再把策略投影到由Q诱导的分布；神经网络版本使用随机重参数化actor、双Q取小值及独立软状态价值网络与目标网络，并用tanh约束…", "abstractZh": "原版SAC将奖励与策略熵共同优化，以离策略回放提高利用率。软策略迭代先评估软Q值，再把策略投影到由Q诱导的分布；神经网络版本使用随机重参数化actor、双Q取小值及独立软状态价值网络与目标网络，并用tanh约束连续动作。\n论文比较Gym的Hopper、Walker2d、HalfCheetah、Ant、Humanoid以及rllab Humanoid，共六项连续控制仿真。每算法五随机种子，每1000环境步评一次展开；曲线阴影是最小至最大而非置信区间。基线含DDPG、PPO、SQL和同期TD3，另消融确定性策略、奖励缩放及目标更新速度。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "sim", "robots": [], "codeStatus": "open", "status": "作者TensorFlow实现公开。", "trainingNote": "回放池训练随机策略与价值网络；此条目对应ICML 2018版本。", "whyUseful": "复现应锁定2018算法版本、独立V网络、动作对数概率的tanh雅可比修正、百万回放池及256批量。作者奖励缩放在多数任务为5、Gym Humanoid为20、rllab为10；不要无说明替换成现代SAC实现后声称复现原结果。", "contribution": "原版SAC将奖励与策略熵共同优化，以离策略回放提高利用率。软策略迭代先评估软Q值，再把策略投影到由Q诱导的分布；神经网络版本使用随机重参数化actor、双Q取小值及独立软状态价值网络与目标网络，并用tanh约束连续动作。", "limitations": "收敛证明针对表格化软策略迭代及相应假设，不等于深度网络全局收敛保证。原版对奖励缩放明显敏感，过小会近均匀、过大易过早确定化；本篇没有实体机器人实验，也没有后续版本的自动温度调节验证。", "caveats": "不要把后续《SAC Algorithms and Applications》的真机实验归给本篇。", "evidence": [ { "url": "https://arxiv.org/abs/1801.01290", "note": "首发日期、作者与最大熵目标。" }, { "url": "https://arxiv.org/pdf/1801.01290", "note": "原论文连续控制仿真评测。" }, { "url": "https://github.com/haarnoja/sac", "note": "作者仓库明确对应ICML 2018论文及TensorFlow实现。" }, { "url": "https://arxiv.org/pdf/1801.01290v2", "note": "4.1：Convergence analysis is tabular soft policy iteration;deep continuous method is approximate." }, { "url": "https://arxiv.org/pdf/1801.01290v2", "note": "4.2/附录C：Original implementation uses separate V,dualQ,reparameterized policy,and tanh density correction." }, { "url": "https://arxiv.org/pdf/1801.01290v2", "note": "5.1/5.2：5seeds;one evaluation every1000steps;shadingmin/max;mean-action evaluation." }, { "url": "https://arxiv.org/pdf/1801.01290v2", "note": "附录D 表1/2：Reward scaling5/20/10 by task;replay1e6,batch256,2×256MLP;no real hardware." } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "基础理论", "强化学习" ], "robotFilters": [], "verificationNote": "2026-10-04核对原论文与作者仓库；区分SAC后续扩展版本。", "original": { "id": "arxiv-1801.01290", "originalSourceUrl": "https://arxiv.org/pdf/1801.01290v2", "versionedOriginalUrl": "https://arxiv.org/pdf/1801.01290v2", "finalDownloadUrl": "https://arxiv.org/pdf/1801.01290", "sha256": "5c33fae017d02f7025730f05198d4a6b103402822c8bbf48cbc5d8a0474c336a", "bytes": 4388102, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:16.963899+00:00", "publicationDate": "2018/01/04", "publicationDateSourceUrl": "https://arxiv.org/abs/1801.01290", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1801.01290", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1801.01290", "pages": 14, "sourceIdentityStatus": "title_match", "sourceTitle": "Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor", "sourceVersion": "1801.01290v2", "sourceVersionDate": "2018/08/08", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.471394+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1801.01290" }, "originalAnalysis": { "id": "arxiv-1801.01290", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739412+00:00", "sourceUrl": "https://arxiv.org/pdf/1801.01290v2", "sectionsRead": [ "3 最大熵目标", "4.1/4.2 软策略迭代与SAC", "5.1/5.2 图1–3评测和消融", "附录C/D 动作压缩与超参" ], "methodsZh": "原版SAC将奖励与策略熵共同优化，以离策略回放提高利用率。软策略迭代先评估软Q值，再把策略投影到由Q诱导的分布；神经网络版本使用随机重参数化actor、双Q取小值及独立软状态价值网络与目标网络，并用tanh约束连续动作。", "experimentsZh": "论文比较Gym的Hopper、Walker2d、HalfCheetah、Ant、Humanoid以及rllab Humanoid，共六项连续控制仿真。每算法五随机种子，每1000环境步评一次展开；曲线阴影是最小至最大而非置信区间。基线含DDPG、PPO、SQL和同期TD3，另消融确定性策略、奖励缩放及目标更新速度。", "resultsZh": "作者报告简单任务表现可比，困难Humanoid类任务学习更快、跨种子较稳定；确定性变体方差更高。评估用策略均值动作往往提高回报，但曲线只计环境奖励，不包含训练目标中的熵，因此需区分训练目标与测量指标。", "limitationsZh": "收敛证明针对表格化软策略迭代及相应假设，不等于深度网络全局收敛保证。原版对奖励缩放明显敏感，过小会近均匀、过大易过早确定化；本篇没有实体机器人实验，也没有后续版本的自动温度调节验证。", "reproductionZh": "复现应锁定2018算法版本、独立V网络、动作对数概率的tanh雅可比修正、百万回放池及256批量。作者奖励缩放在多数任务为5、Gym Humanoid为20、rllab为10；不要无说明替换成现代SAC实现后声称复现原结果。", "evidenceNotes": [ { "section/page": "4.1", "note": "Convergence analysis is tabular soft policy iteration;deep continuous method is approximate." }, { "section/page": "4.2/附录C", "note": "Original implementation uses separate V,dualQ,reparameterized policy,and tanh density correction." }, { "section/page": "5.1/5.2", "note": "5seeds;one evaluation every1000steps;shadingmin/max;mean-action evaluation." }, { "section/page": "附录D 表1/2", "note": "Reward scaling5/20/10 by task;replay1e6,batch256,2×256MLP;no real hardware." } ], "experimentType": "sim", "robots": [], "corrections": [ "勿把自动温度或SAC后续机器人论文的实机实验归于本篇。" ], "sourceVersion": "1801.01290v2", "originalSha256": "5c33fae017d02f7025730f05198d4a6b103402822c8bbf48cbc5d8a0474c336a" }, "experimentNote": "论文比较Gym的Hopper、Walker2d、HalfCheetah、Ant、Humanoid以及rllab Humanoid，共六项连续控制仿真。每算法五随机种子，每1000环境步评一次展开；曲线阴影是最小至最大而非置信区间。基线含DDPG、PPO、SQL和同期TD3，另消融确定性策略、奖励缩放及目标更新速度。", "analysisVerifiedAt": "2026-10-04T13:47:58.739412+00:00" }, { "id": "arxiv-1711.07280", "title": "Vision-and-Language Navigation: Interpreting visually-grounded navigation instructions in real environments", "titleZh": "视觉与语言导航：理解真实环境中的视觉落地导航指令", "date": "2017-11-20", "datePrecision": "day", "year": 2017, "url": "https://arxiv.org/abs/1711.07280", "paperUrl": "https://arxiv.org/abs/1711.07280", "codeUrl": "https://github.com/peteanderson80/Matterport3DSimulator", "summary": "提出基于Matterport真实全景的离散导航模拟器与R2R语言路线基准。基线用LSTM编码指令，注意力解码器结合当前ResNet图像特征和上一动作；比较教师强制与学生自行采样行动，再用当前位置到目标最短路监督…", "abstractZh": "提出基于Matterport真实全景的离散导航模拟器与R2R语言路线基准。基线用LSTM编码指令，注意力解码器结合当前ResNet图像特征和上一动作；比较教师强制与学生自行采样行动，再用当前位置到目标最短路监督。\n数据含21567条指令、平均29词、每路径三描述；61建筑用于训练/已见验证，11未见验证，18测试。智能体六种动作含停止，最终距目标小于3米算成功，另报沿轨迹最佳停止的oracle成功；1390测试指令有人类参照。", "category": "具身导航 / 视觉语言基准", "tags": [ "VLN", "R2R", "Matterport3D", "语言导航" ], "directions": [ "数据集与基准", "导航与建图" ], "tier": "foundation", "experimentType": "sim", "robots": [], "robotFilters": [], "codeStatus": "open", "status": "实现已开源", "trainingNote": "官方仓库提供模拟器、R2R 数据准备和学习基线；场景数据需遵循独立 Matterport3D 使用条款。", "contribution": "提出基于Matterport真实全景的离散导航模拟器与R2R语言路线基准。基线用LSTM编码指令，注意力解码器结合当前ResNet图像特征和上一动作；比较教师强制与学生自行采样行动，再用当前位置到目标最短路监督。", "whyUseful": "复现须严格按建筑划分、固定视场和30度相机转动、词表与训练策略，测试提交模型使用训练及验证数据的设置应透明。模拟器代码与Matterport场景授权分别处理，本轮未运行基线或重新标注数据。", "limitations": "真实图像来源不等于实体机器人；动作沿预定义导航图跳转，弱化避障、动力学和视觉里程计。终点3米指标不检查完整路线遵从，oracle成功更不能视作机器人自行判断完成；没有实机型号。", "license": "MIT（模拟器代码）；场景数据另受 Matterport3D 条款约束", "evidence": [ { "url": "https://arxiv.org/abs/1711.07280", "note": "首版日期、R2R 与基于真实图像的模拟器任务范围。" }, { "url": "https://github.com/peteanderson80/Matterport3DSimulator", "note": "作者模拟器与 R2R 基线入口。" }, { "url": "https://github.com/peteanderson80/Matterport3DSimulator/blob/master/LICENSE", "note": "源码 LICENSE 为 MIT；不替代场景数据授权。" }, { "url": "https://arxiv.org/pdf/1711.07280", "note": "§4.4：最终3米内并选择停止；oracle是另一个指标。" }, { "url": "https://arxiv.org/pdf/1711.07280", "note": "Table 1：测试20.4%对人类86.4%；环境泛化差距。" }, { "url": "https://arxiv.org/pdf/1711.07280", "note": "§3 / §5：预定义视点图与六动作；无实体机器人。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "以 R2R 将自然语言路线理解变成具身导航的标准化评测问题。", "experimentNote": "数据含21567条指令、平均29词、每路径三描述；61建筑用于训练/已见验证，11未见验证，18测试。智能体六种动作含停止，最终距目标小于3米算成功，另报沿轨迹最佳停止的oracle成功；1390测试指令有人类参照。", "projectUrl": null, "shortTitle": "", "original": { "id": "arxiv-1711.07280", "originalSourceUrl": "https://arxiv.org/pdf/1711.07280v3", "versionedOriginalUrl": "https://arxiv.org/pdf/1711.07280v3", "finalDownloadUrl": "https://arxiv.org/pdf/1711.07280", "sha256": "b7589865ee573d4753ee9e25bdc494cbbb199e6ec17d0a349091aed0fd86c420", "bytes": 9720283, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:16.979569+00:00", "publicationDate": "2017/11/20", "publicationDateSourceUrl": "https://arxiv.org/abs/1711.07280", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1711.07280", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1711.07280", "pages": 13, "sourceIdentityStatus": "title_match", "sourceTitle": "Vision-and-Language Navigation: Interpreting visually-grounded navigation instructions in real environments", "sourceVersion": "1711.07280v3", "sourceVersionDate": "2018/04/05", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.481156+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1711.07280" }, "originalAnalysis": { "id": "arxiv-1711.07280", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090152+00:00", "sourceUrl": "https://arxiv.org/pdf/1711.07280", "sectionsRead": [ "§3", "§4.1–4.4", "§5.1–5.3", "Table 1", "§6" ], "methodsZh": "提出基于Matterport真实全景的离散导航模拟器与R2R语言路线基准。基线用LSTM编码指令，注意力解码器结合当前ResNet图像特征和上一动作；比较教师强制与学生自行采样行动，再用当前位置到目标最短路监督。", "experimentsZh": "数据含21567条指令、平均29词、每路径三描述；61建筑用于训练/已见验证，11未见验证，18测试。智能体六种动作含停止，最终距目标小于3米算成功，另报沿轨迹最佳停止的oracle成功；1390测试指令有人类参照。", "resultsZh": "作者表1学生强制在已见验证38.6%、未见验证21.8%、测试20.4%，测试人类86.4%、随机13.2%。明显的已见/未见差距使环境泛化成为关键问题，而正确停止本身也占据显著难度。", "limitationsZh": "真实图像来源不等于实体机器人；动作沿预定义导航图跳转，弱化避障、动力学和视觉里程计。终点3米指标不检查完整路线遵从，oracle成功更不能视作机器人自行判断完成；没有实机型号。", "reproductionZh": "复现须严格按建筑划分、固定视场和30度相机转动、词表与训练策略，测试提交模型使用训练及验证数据的设置应透明。模拟器代码与Matterport场景授权分别处理，本轮未运行基线或重新标注数据。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "§4.4", "note": "最终3米内并选择停止；oracle是另一个指标。" }, { "section/page": "Table 1", "note": "测试20.4%对人类86.4%；环境泛化差距。" }, { "section/page": "§3 / §5", "note": "预定义视点图与六动作；无实体机器人。" } ] }, "analysisVerifiedAt": "2026-10-04T13:50:35.090152+00:00" }, { "id": "arxiv-1708.03852", "title": "VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator", "titleZh": "VINS-Mono：稳健通用的单目视觉惯性状态估计器", "date": "2017-08-13", "datePrecision": "day", "year": 2017, "url": "https://arxiv.org/abs/1708.03852", "paperUrl": "https://arxiv.org/abs/1708.03852", "codeUrl": "https://github.com/HKUST-Aerial-Robotics/VINS-Mono", "summary": "先将视觉SfM与IMU预积分对齐初始化尺度、重力、速度和陀螺偏置，再在滑窗中紧耦合优化视觉残差、惯性残差及边缘化先验。回环重定位提供跨帧约束，四自由度位姿图修正位置和偏航漂移，并包含失效检测与重启。", "abstractZh": "先将视觉SfM与IMU预积分对齐初始化尺度、重力、速度和陀螺偏置，再在滑窗中紧耦合优化视觉残差、惯性残差及边缘化先验。回环重定位提供跨帧约束，四自由度位姿图修正位置和偏航漂移，并包含失效检测与重启。\nEuRoC使用左相机与IMU比较OKVIS，另做手持室内/室外及5.62公里校园序列。真实四旋翼用机载状态估计闭环跟踪四轮八字轨迹，关闭回环并以OptiTrack测真值；还在iPhone7 Plus上展示移动AR。", "category": "视觉惯性 / 状态估计", "tags": [ "VINS-Mono", "IMU 预积分", "紧耦合", "视觉惯性里程计" ], "directions": [ "其他机器人研究" ], "tier": "classic", "experimentType": "real", "robots": [ "四旋翼（完整型号未注明）" ], "robotFilters": [], "codeStatus": "open", "status": "实现已开源", "trainingNote": "无需神经网络训练；官方 GPLv3 仓库包含估计器、特征跟踪、位姿图及运行配置。", "contribution": "先将视觉SfM与IMU预积分对齐初始化尺度、重力、速度和陀螺偏置，再在滑窗中紧耦合优化视觉残差、惯性残差及边缘化先验。回环重定位提供跨帧约束，四自由度位姿图修正位置和偏航漂移，并包含失效检测与重启。", "whyUseful": "应记录相机模型、IMU噪声/外参、启动运动和回环开关，区分VIO估计与闭环飞行误差。先用公开序列验证时间同步和尺度，再做有保护的控制测试；本次未运行代码。", "limitations": "单目初始化需要足够视差与惯性激励，标定、同步和IMU偏置很关键。回环改变的是长期一致性，不保证每时刻误差；室外长程演示不能与带全程真值的短序列直接比较。四旋翼完整机型未明确核实。", "license": "GPL-3.0", "evidence": [ { "url": "https://arxiv.org/abs/1708.03852", "note": "作者摘要和首版日期；明确公开数据、实机闭环飞行与 iOS 演示。" }, { "url": "https://github.com/HKUST-Aerial-Robotics/VINS-Mono", "note": "官方仓库核实 GPLv3 和完整模块；README 区分后续增补功能。" }, { "url": "https://arxiv.org/pdf/1708.03852", "note": "§V–VIII：SfM惯性初始化、滑窗优化及4DoF位姿图。" }, { "url": "https://arxiv.org/pdf/1708.03852", "note": "§IX-C–E：5.62公里缺完整真值；61.97米闭环飞行0.29%漂移。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "将视觉惯性的关键数学模块连成可用于真实闭环飞行的系统化开源方案。", "experimentNote": "EuRoC使用左相机与IMU比较OKVIS，另做手持室内/室外及5.62公里校园序列。真实四旋翼用机载状态估计闭环跟踪四轮八字轨迹，关闭回环并以OptiTrack测真值；还在iPhone7 Plus上展示移动AR。", "projectUrl": null, "shortTitle": "", "original": { "id": "arxiv-1708.03852", "originalSourceUrl": "https://arxiv.org/pdf/1708.03852v1", "versionedOriginalUrl": "https://arxiv.org/pdf/1708.03852v1", "finalDownloadUrl": "https://arxiv.org/pdf/1708.03852", "sha256": "1f44f92f057cbb05a6570f14d2fe5cae2b9736198a8eeca7c93438adbc38a087", "bytes": 4382711, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:27.278976+00:00", "publicationDate": "2017/08/13", "publicationDateSourceUrl": "https://arxiv.org/abs/1708.03852", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1708.03852", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1708.03852", "pages": 17, "sourceIdentityStatus": "title_match", "sourceTitle": "VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator", "sourceVersion": "1708.03852v1", "sourceVersionDate": "2017/08/13", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.485710+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1708.03852" }, "originalAnalysis": { "id": "arxiv-1708.03852", "sourceUrl": "https://arxiv.org/pdf/1708.03852", "sectionsRead": [ "§IV–VI", "§VII–VIII", "§IX-A–E", "Table I" ], "methodsZh": "先将视觉SfM与IMU预积分对齐初始化尺度、重力、速度和陀螺偏置，再在滑窗中紧耦合优化视觉残差、惯性残差及边缘化先验。回环重定位提供跨帧约束，四自由度位姿图修正位置和偏航漂移，并包含失效检测与重启。", "experimentsZh": "EuRoC使用左相机与IMU比较OKVIS，另做手持室内/室外及5.62公里校园序列。真实四旋翼用机载状态估计闭环跟踪四轮八字轨迹，关闭回环并以OptiTrack测真值；还在iPhone7 Plus上展示移动AR。", "resultsZh": "作者报告61.97米飞行末端漂移约0.29%，100Hz状态反馈支持控制。EuRoC比较显示纯VIO与OKVIS各有优势，完整回环系统改善位置漂移；大型校园轨迹主要与地图视觉对齐，没有全程高精度真值，不能当作厘米级准确率。", "limitationsZh": "单目初始化需要足够视差与惯性激励，标定、同步和IMU偏置很关键。回环改变的是长期一致性，不保证每时刻误差；室外长程演示不能与带全程真值的短序列直接比较。四旋翼完整机型未明确核实。", "reproductionZh": "应记录相机模型、IMU噪声/外参、启动运动和回环开关，区分VIO估计与闭环飞行误差。先用公开序列验证时间同步和尺度，再做有保护的控制测试；本次未运行代码。", "experimentType": "real", "robots": [ "四旋翼（完整型号未注明）" ], "evidenceNotes": [ { "section/page": "§V–VIII", "note": "SfM惯性初始化、滑窗优化及4DoF位姿图。" }, { "section/page": "§IX-C–E", "note": "5.62公里缺完整真值；61.97米闭环飞行0.29%漂移。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-1707.06347", "title": "Proximal Policy Optimization Algorithms", "titleZh": "近端策略优化算法", "shortTitle": "PPO", "date": "2017-07-20", "datePrecision": "day", "year": 2017, "url": "https://arxiv.org/abs/1707.06347", "codeUrl": "https://github.com/openai/baselines", "category": "强化学习", "tags": [ "策略梯度", "同策略学习", "基础算法" ], "tier": "foundation", "summary": "PPO收集当前策略批次后，用旧/新动作概率比和优势估计构造替代目标，多轮小批量更新。裁剪版本取原目标与裁剪目标较小者，移除向过大改善方向继续推进的收益；另一版本自适应调节KL惩罚，并结合价值和熵目标。", "abstractZh": "PPO收集当前策略批次后，用旧/新动作概率比和优势估计构造替代目标，多轮小批量更新。裁剪版本取原目标与裁剪目标较小者，移除向过大改善方向继续推进的收益；另一版本自适应调节KL惩罚，并结合价值和熵目标。\n七个MuJoCo任务各一百万步比较替代目标，再与TRPO、A2C、CEM等比连续控制；Roboschool模拟人形奔跑/转向/受击起身作为扩展，49个Atari游戏比较学习过程及最终得分，三次试验汇总。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "sim", "robots": [], "codeStatus": "open", "status": "作者机构OpenAI Baselines公开PPO实现；依赖较旧。", "trainingNote": "On-policy采样后做多轮更新；工程细节影响结果。", "whyUseful": "先固定优势估计、概率比、裁剪范围0.2、epoch和minibatch，保存旧策略log概率及终止掩码；同时监测KL、裁剪比例、熵和多个随机种子。比较采样效率与最终成绩须使用相同环境步数。", "contribution": "PPO收集当前策略批次后，用旧/新动作概率比和优势估计构造替代目标，多轮小批量更新。裁剪版本取原目标与裁剪目标较小者，移除向过大改善方向继续推进的收益；另一版本自适应调节KL惩罚，并结合价值和熵目标。", "limitations": "裁剪目标不是硬KL约束，更不是机器人安全保证；依旧同策略，需要新环境采样，不能任意无限复用旧数据。模拟人形和Atari不构成实机证据，超参数及优势、价值实现会影响结果。", "caveats": "原论文没有实体机器人验证；后续PPO机器人应用需分别阅读。", "evidence": [ { "url": "https://arxiv.org/abs/1707.06347", "note": "首发日期、替代目标与模拟运动/Atari评测。" }, { "url": "https://github.com/openai/baselines", "note": "官方机构实现、MIT许可、PPO1/PPO2及历史TensorFlow依赖。" }, { "url": "https://arxiv.org/pdf/1707.06347", "note": "3–5：裁剪/KL两版本及多epoch更新。" }, { "url": "https://arxiv.org/pdf/1707.06347", "note": "6.1–6.4 Tables 1–2：七连续任务、49游戏及过程/最终指标差异。" }, { "url": "https://arxiv.org/pdf/1707.06347v2", "note": "3–5：裁剪/KL两版本及多epoch更新。" }, { "url": "https://arxiv.org/pdf/1707.06347v2", "note": "6.1–6.4 Tables 1–2：七连续任务、49游戏及过程/最终指标差异。" } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "基础理论", "强化学习" ], "robotFilters": [], "verificationNote": "2026-10-04核对论文与官方实现；代码公开不等于兼容当前依赖。", "original": { "id": "arxiv-1707.06347", "originalSourceUrl": "https://arxiv.org/pdf/1707.06347v2", "versionedOriginalUrl": "https://arxiv.org/pdf/1707.06347v2", "finalDownloadUrl": "https://arxiv.org/pdf/1707.06347", "sha256": "e78feadadbdbb0b601b3c2bcc81404722cd431a489b307545f9b7bea1e8c4f5b", "bytes": 2923532, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:27.291864+00:00", "publicationDate": "2017/07/20", "publicationDateSourceUrl": "https://arxiv.org/abs/1707.06347", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1707.06347", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1707.06347", "pages": 12, "sourceIdentityStatus": "title_match", "sourceTitle": "Proximal Policy Optimization Algorithms", "sourceVersion": "1707.06347v2", "sourceVersionDate": "2017/08/28", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.489201+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1707.06347" }, "originalAnalysis": { "id": "arxiv-1707.06347", "sourceUrl": "https://arxiv.org/pdf/1707.06347v2", "sectionsRead": [ "2", "3", "4", "5", "6.1–6.4", "7", "Tables 1–2" ], "methodsZh": "PPO收集当前策略批次后，用旧/新动作概率比和优势估计构造替代目标，多轮小批量更新。裁剪版本取原目标与裁剪目标较小者，移除向过大改善方向继续推进的收益；另一版本自适应调节KL惩罚，并结合价值和熵目标。", "experimentsZh": "七个MuJoCo任务各一百万步比较替代目标，再与TRPO、A2C、CEM等比连续控制；Roboschool模拟人形奔跑/转向/受击起身作为扩展，49个Atari游戏比较学习过程及最终得分，三次试验汇总。", "resultsZh": "裁剪目标通常优于固定或自适应KL方案，连续控制综合表现强；Atari按全程平均奖励PPO赢30/49个游戏、ACER18个，但按最后100局ACER赢28个、PPO19个。优势是实现、稳定性与样本复用折中，并非所有指标始终最优。", "limitationsZh": "裁剪目标不是硬KL约束，更不是机器人安全保证；依旧同策略，需要新环境采样，不能任意无限复用旧数据。模拟人形和Atari不构成实机证据，超参数及优势、价值实现会影响结果。", "reproductionZh": "先固定优势估计、概率比、裁剪范围0.2、epoch和minibatch，保存旧策略log概率及终止掩码；同时监测KL、裁剪比例、熵和多个随机种子。比较采样效率与最终成绩须使用相同环境步数。", "experimentType": "sim", "robots": [], "corrections": [ "PPO在Atari全程学习指标优胜，但最终得分获胜游戏数少于ACER。" ], "evidenceNotes": [ { "note": "裁剪/KL两版本及多epoch更新。", "section": "3–5" }, { "note": "七连续任务、49游戏及过程/最终指标差异。", "section": "6.1–6.4 Tables 1–2" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "1707.06347v2", "originalSha256": "e78feadadbdbb0b601b3c2bcc81404722cd431a489b307545f9b7bea1e8c4f5b" }, "experimentNote": "七个MuJoCo任务各一百万步比较替代目标，再与TRPO、A2C、CEM等比连续控制；Roboschool模拟人形奔跑/转向/受击起身作为扩展，49个Atari游戏比较学习过程及最终得分，三次试验汇总。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-1703.06907", "title": "Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World", "titleZh": "通过域随机化将深度网络从仿真迁移到现实", "shortTitle": "Domain Randomization", "date": "2017-03-20", "datePrecision": "day", "year": 2017, "url": "https://arxiv.org/abs/1703.06907", "codeUrl": null, "category": "Sim2Real", "tags": [ "视觉随机化", "物体定位", "抓取" ], "tier": "classic", "summary": "论文随机化合成场景纹理、灯光、物体位置、干扰物、相机姿态与视场，训练VGG式卷积网络从单幅RGB回归目标中心坐标。已知目标形状尺寸、固定桌高降低了几何难度；无需逼真纹理或真实图像微调，模型学到的是位置估计，不是…", "abstractZh": "论文随机化合成场景纹理、灯光、物体位置、干扰物、相机姿态与视场，训练VGG式卷积网络从单幅RGB回归目标中心坐标。已知目标形状尺寸、固定桌高降低了几何难度；无需逼真纹理或真实图像微调，模型学到的是位置估计，不是完整抓取控制。\n八个几何物体各有60张真实测试图，分独物体、干扰和遮挡三组，共480张；相机距离约70至105厘米。比较合成图数量、纹理数量、ImageNet初始化、噪声与相机/干扰物随机化；最后将定位器接常规运动规划和预设抓取，在Fetch上测试。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "Fetch" ], "codeStatus": "unknown", "status": "本次未核实原论文官方代码。", "trainingNote": "合成图像监督训练定位器；真实抓取用于验证。", "whyUseful": "复现需物体网格、MuJoCo渲染、随机化范围及真实网格测位协议，单独核对抓取规划器。应固定最佳模型选择方式并逐类别报告误差和40/10次试验；区分视觉合成训练与整个机器人系统是否经过标定或人工设计。", "contribution": "论文随机化合成场景纹理、灯光、物体位置、干扰物、相机姿态与视场，训练VGG式卷积网络从单幅RGB回归目标中心坐标。已知目标形状尺寸、固定桌高降低了几何难度；无需逼真纹理或真实图像微调，模型学到的是位置估计，不是完整抓取控制。", "limitations": "低样本、高度固定、简单几何以及选择表现较好的两种检测器限制了外推；相机只在近似指定范围随机变化，不能理解为任意视角免校准。真实定位误差仍大于仿真，尚未验证高精度接触操作或动力学策略迁移。", "caveats": "定位与抓取是分开的系统步骤，不应称为纯仿真训练的完整通用策略。", "evidence": [ { "url": "https://arxiv.org/abs/1703.06907", "note": "首发日期、仿真RGB训练与真实定位摘要。" }, { "url": "https://arxiv.org/pdf/1703.06907", "note": "真实抓取部分明确Fetch机器人和完整流水线。" }, { "url": "https://arxiv.org/pdf/1703.06907v1", "note": "III-A/B：Fixed table height;camera random within10×5×10cm box,angles±0.1rad,FOV±5%;VGG16-style position regression." }, { "url": "https://arxiv.org/pdf/1703.06907v1", "note": "IV-B 表I：480 real images;8objects×3conditions×20images;~1.5cm mean real localization." }, { "url": "https://arxiv.org/pdf/1703.06907v1", "note": "IV-C 表II：20k training samples;distractor error1.8→7.2cm without distractors." }, { "url": "https://arxiv.org/pdf/1703.06907v1", "note": "IV-D：Fetch;prescribed grasp+off-the-shelf motion planner;38/40 selected geometric-object trials;9/10Spam." } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "操作与抓取" ], "robotFilters": [ "Fetch" ], "verificationNote": "2026-10-04核对原论文；未将视觉随机化与后续动力学随机化论文混同。", "original": { "id": "arxiv-1703.06907", "originalSourceUrl": "https://arxiv.org/pdf/1703.06907v1", "versionedOriginalUrl": "https://arxiv.org/pdf/1703.06907v1", "finalDownloadUrl": "https://arxiv.org/pdf/1703.06907", "sha256": "3fc98c5f4cea050686d45858e647e1b704492121fdae80f8fcd5d560c107f11f", "bytes": 4066798, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:31.581183+00:00", "publicationDate": "2017/03/20", "publicationDateSourceUrl": "https://arxiv.org/abs/1703.06907", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1703.06907", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1703.06907", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World", "sourceVersion": "1703.06907v1", "sourceVersionDate": "2017/03/20", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.495176+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1703.06907" }, "originalAnalysis": { "id": "arxiv-1703.06907", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739414+00:00", "sourceUrl": "https://arxiv.org/pdf/1703.06907v1", "sectionsRead": [ "III 域随机化、VGG式位置回归", "IV-A/B 数据与表I定位", "IV-C 图4/5和表II消融", "IV-D Fetch抓取及V局限" ], "methodsZh": "论文随机化合成场景纹理、灯光、物体位置、干扰物、相机姿态与视场，训练VGG式卷积网络从单幅RGB回归目标中心坐标。已知目标形状尺寸、固定桌高降低了几何难度；无需逼真纹理或真实图像微调，模型学到的是位置估计，不是完整抓取控制。", "experimentsZh": "八个几何物体各有60张真实测试图，分独物体、干扰和遮挡三组，共480张；相机距离约70至105厘米。比较合成图数量、纹理数量、ImageNet初始化、噪声与相机/干扰物随机化；最后将定位器接常规运动规划和预设抓取，在Fetch上测试。", "resultsZh": "作者报告真实平均定位误差约1.5厘米，仿真为0.3至0.5厘米。两种较稳定目标在40次杂乱抓取中成功38次，另对Spam罐10次成功9次。表II固定两万训练图时，去干扰物随机化使干扰测试误差由1.8厘米增至7.2厘米；加普通图像噪声作用较小。", "limitationsZh": "低样本、高度固定、简单几何以及选择表现较好的两种检测器限制了外推；相机只在近似指定范围随机变化，不能理解为任意视角免校准。真实定位误差仍大于仿真，尚未验证高精度接触操作或动力学策略迁移。", "reproductionZh": "复现需物体网格、MuJoCo渲染、随机化范围及真实网格测位协议，单独核对抓取规划器。应固定最佳模型选择方式并逐类别报告误差和40/10次试验；区分视觉合成训练与整个机器人系统是否经过标定或人工设计。", "evidenceNotes": [ { "section/page": "III-A/B", "note": "Fixed table height;camera random within10×5×10cm box,angles±0.1rad,FOV±5%;VGG16-style position regression." }, { "section/page": "IV-B 表I", "note": "480 real images;8objects×3conditions×20images;~1.5cm mean real localization." }, { "section/page": "IV-C 表II", "note": "20k training samples;distractor error1.8→7.2cm without distractors." }, { "section/page": "IV-D", "note": "Fetch;prescribed grasp+off-the-shelf motion planner;38/40 selected geometric-object trials;9/10Spam." } ], "experimentType": "both", "robots": [ "Fetch" ], "corrections": [ "仅视觉定位模型从合成数据迁移，抓取由规划器和预设动作完成。" ], "sourceVersion": "1703.06907v1", "originalSha256": "3fc98c5f4cea050686d45858e647e1b704492121fdae80f8fcd5d560c107f11f" }, "experimentNote": "八个几何物体各有60张真实测试图，分独物体、干扰和遮挡三组，共480张；相机距离约70至105厘米。比较合成图数量、纹理数量、ImageNet初始化、噪声与相机/干扰物随机化；最后将定位器接常规运动规划和预设抓取，在Fetch上测试。", "analysisVerifiedAt": "2026-10-04T13:47:58.739414+00:00" }, { "id": "arxiv-1610.06475", "title": "ORB-SLAM2: an Open-Source SLAM System for Monocular, Stereo and RGB-D Cameras", "titleZh": "ORB-SLAM2：支持单目、双目和 RGB-D 相机的开源 SLAM 系统", "date": "2016-10-20", "datePrecision": "day", "year": 2016, "url": "https://arxiv.org/abs/1610.06475", "paperUrl": "https://arxiv.org/abs/1610.06475", "codeUrl": "https://github.com/raulmur/ORB_SLAM2", "summary": "ORB-SLAM2把单目、双目与RGB-D观测统一到ORB特征跟踪、局部束调整和回环框架；深度转成虚拟右目坐标，近点提供尺度/平移约束，远点帮助旋转。维护共视图与生成树进行地图复用、重定位、位姿图和全局束调整。", "abstractZh": "ORB-SLAM2把单目、双目与RGB-D观测统一到ORB特征跟踪、局部束调整和回环框架；深度转成虚拟右目坐标，近点提供尺度/平移约束，远点帮助旋转。维护共视图与生成树进行地图复用、重定位、位姿图和全局束调整。\n在真实传感器序列KITTI、EuRoC、TUM RGB-D测试，Intel i7-4790与16GB内存，每序列运行五次取精度中位数。与作者发表的同期系统结果比较，报告绝对/相对误差和各线程耗时；实验是数据序列评估而非机器人导航闭环。", "category": "视觉 SLAM / 多相机配置", "tags": [ "ORB-SLAM2", "双目", "RGB-D", "回环检测", "真实数据评测" ], "directions": [ "导航与建图" ], "tier": "classic", "experimentType": "data", "robots": [], "robotFilters": [], "codeStatus": "open", "status": "实现已开源", "trainingNote": "无需神经策略训练；作者仓库提供 KITTI、TUM RGB-D、EuRoC 配置和 ROS/非 ROS 运行入口。", "contribution": "ORB-SLAM2把单目、双目与RGB-D观测统一到ORB特征跟踪、局部束调整和回环框架；深度转成虚拟右目坐标，近点提供尺度/平移约束，远点帮助旋转。维护共视图与生成树进行地图复用、重定位、位姿图和全局束调整。", "whyUseful": "复现应使用随源码给出的标定、词袋和配置，说明是否补偿深度尺度、是否启用回环及局部化模式，记录五次中位数和失败段。无需神经训练；代码、场景数据及ROS依赖需分别核验，本轮未重跑。", "limitations": "EuRoC V2_03严重模糊时部分丢跟踪；依赖纹理、标定与场景条件。TUM freiburg2深度存在约4%尺度偏差，本文进行补偿，可能贡献对其他公布结果的优势。稀疏定位地图也不等于完整可通行地图。", "license": "GPL-3.0", "evidence": [ { "url": "https://arxiv.org/abs/1610.06475", "note": "arXiv 确认首版日期、29 段基准和多输入系统范围。" }, { "url": "https://github.com/raulmur/ORB_SLAM2", "note": "作者仓库确认 GPLv3、真实尺度适用配置和数据集运行示例。" }, { "url": "https://arxiv.org/pdf/1610.06475", "note": "§IV：五次运行精度中位数；真实序列，不是实机导航。" }, { "url": "https://arxiv.org/pdf/1610.06475", "note": "§IV-C/Table III：freiburg2补偿约4%深度尺度偏差。" }, { "url": "https://arxiv.org/pdf/1610.06475", "note": "§IV-B：V2_03运动模糊导致部分丢失。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "把成熟特征型 SLAM 扩展为可统一运行单目、双目和 RGB-D 的开源基线。", "experimentNote": "在真实传感器序列KITTI、EuRoC、TUM RGB-D测试，Intel i7-4790与16GB内存，每序列运行五次取精度中位数。与作者发表的同期系统结果比较，报告绝对/相对误差和各线程耗时；实验是数据序列评估而非机器人导航闭环。", "projectUrl": null, "shortTitle": "", "original": { "id": "arxiv-1610.06475", "originalSourceUrl": "https://arxiv.org/pdf/1610.06475v2", "versionedOriginalUrl": "https://arxiv.org/pdf/1610.06475v2", "finalDownloadUrl": "https://arxiv.org/pdf/1610.06475", "sha256": "2b94552a20378d6dfbe863ba7317d4a86874fe54e43ca8b918abefcbdc64ce24", "bytes": 4187799, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:34.914844+00:00", "publicationDate": "2016/10/20", "publicationDateSourceUrl": "https://arxiv.org/abs/1610.06475", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1610.06475", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1610.06475", "pages": 9, "sourceIdentityStatus": "title_match", "sourceTitle": "ORB-SLAM2: an Open-Source SLAM System for Monocular, Stereo and RGB-D Cameras", "sourceVersion": "1610.06475v2", "sourceVersionDate": "2017/06/19", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.501740+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1610.06475" }, "originalAnalysis": { "id": "arxiv-1610.06475", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090153+00:00", "sourceUrl": "https://arxiv.org/pdf/1610.06475", "sectionsRead": [ "§III", "§IV-A–D", "Tables I–IV", "§V" ], "methodsZh": "ORB-SLAM2把单目、双目与RGB-D观测统一到ORB特征跟踪、局部束调整和回环框架；深度转成虚拟右目坐标，近点提供尺度/平移约束，远点帮助旋转。维护共视图与生成树进行地图复用、重定位、位姿图和全局束调整。", "experimentsZh": "在真实传感器序列KITTI、EuRoC、TUM RGB-D测试，Intel i7-4790与16GB内存，每序列运行五次取精度中位数。与作者发表的同期系统结果比较，报告绝对/相对误差和各线程耗时；实验是数据序列评估而非机器人导航闭环。", "resultsZh": "作者在多数KITTI序列报告相对误差低于1%，EuRoC多序列达到厘米级；TUM fr1/desk绝对RMSE为0.016米、fr3/office0.010米。深度配置具有真实尺度，单目不能据此自动消除尺度不确定性。", "limitationsZh": "EuRoC V2_03严重模糊时部分丢跟踪；依赖纹理、标定与场景条件。TUM freiburg2深度存在约4%尺度偏差，本文进行补偿，可能贡献对其他公布结果的优势。稀疏定位地图也不等于完整可通行地图。", "reproductionZh": "复现应使用随源码给出的标定、词袋和配置，说明是否补偿深度尺度、是否启用回环及局部化模式，记录五次中位数和失败段。无需神经训练；代码、场景数据及ROS依赖需分别核验，本轮未重跑。", "experimentType": "data", "robots": [], "evidenceNotes": [ { "section/page": "§IV", "note": "五次运行精度中位数；真实序列，不是实机导航。" }, { "section/page": "§IV-C/Table III", "note": "freiburg2补偿约4%深度尺度偏差。" }, { "section/page": "§IV-B", "note": "V2_03运动模糊导致部分丢失。" } ] }, "analysisVerifiedAt": "2026-10-04T13:50:35.090153+00:00" }, { "id": "arxiv-1610.00696", "title": "Deep Visual Foresight for Planning Robot Motion", "titleZh": "用于机器人运动规划的深度视觉预见", "shortTitle": "Visual Foresight", "date": "2016-10-03", "year": 2016, "url": "https://arxiv.org/abs/1610.00696", "paperUrl": "https://arxiv.org/abs/1610.00696", "projectUrl": null, "codeUrl": null, "summary": "动作条件卷积LSTM预测图像变换核与合成掩模，无需物体模型或显式光流监督。把用户指定像素随预测流传播为位置分布，CEM选择使像素接近目标的动作；执行后用真实光流更新像素位置并重新规划。", "abstractZh": "动作条件卷积LSTM预测图像变换核与合成掩模，无需物体模型或显式光流监督。把用户指定像素随预测流传播为位置分布，CEM选择使像素接近目标的动作；执行后用真实光流更新像素位置并重新规划。\n真实7自由度机械臂推动七件未见物体，比较随机动作、末端直接趋向目标及连续视觉伺服。使用64×64图像、三步约800毫秒预测时域，每约200毫秒重规划；训练数据来自十台具有不同相机布置的机器人。", "category": "世界模型 / 视觉规划", "tags": [ "Visual Foresight", "世界模型", "强化学习", "操作与抓取" ], "directions": [ "世界模型", "操作与抓取", "强化学习" ], "tier": "foundation", "experimentType": "real", "experimentNote": "真实7自由度机械臂推动七件未见物体，比较随机动作、末端直接趋向目标及连续视觉伺服。使用64×64图像、三步约800毫秒预测时域，每约200毫秒重规划；训练数据来自十台具有不同相机布置的机器人。", "robots": [ "7自由度机械臂（型号未明确）" ], "robotFilters": [], "robotNote": "原文写明 7-DoF 机械臂；具体厂商和型号未知，未从图片猜测。", "codeStatus": "unknown", "status": "官方代码未核实", "trainingStatus": "官方代码未核实", "trainingNote": "论文描述自监督数据和 MPC 方法；本轮未核实该版本完整官方训练代码。", "codeStatusNote": "本轮未核实可复现该论文的官方训练仓库；不据此断言从未发布。", "license": "未核实", "contribution": "动作条件卷积LSTM预测图像变换核与合成掩模，无需物体模型或显式光流监督。把用户指定像素随预测流传播为位置分布，CEM选择使像素接近目标的动作；执行后用真实光流更新像素位置并重新规划。", "whyUseful": "需复现状态动作同步、视频预测训练、像素目标和光流跟踪，并用同一像素追踪协议比较基线。应同时检查遮挡失败、物体旋转与最终像素距离；本次未补猜硬件或运行代码。", "limitations": "任务集中于短时域反应式推物，候选时域内动作绑定不变，不能推断长期规划能力。机械臂遮挡导致像素跟踪漂到手臂，质量/接触预测偏差会把旋转变成平移；原文所读部分未明确厂商型号。", "caveats": "主要为非抓取推物；规划窗口短，质量和接触误差仍会造成失败。确切机械臂品牌未在核对段落中给出。", "evidence": [ { "url": "https://arxiv.org/abs/1610.00696", "note": "首发日期和标题。" }, { "url": "https://arxiv.org/html/1610.00696v1", "note": "§III、§V-A：7 自由度真实机械臂、视频预测、像素目标及短时域 MPC。" }, { "url": "https://arxiv.org/pdf/1610.00696", "note": "§III–IV; Algorithm 1：动作条件像素流与CEM视觉MPC。" }, { "url": "https://arxiv.org/pdf/1610.00696", "note": "§V-A–C; Table I; Figure 6：800毫秒时域、像素误差、遮挡与物理预测失败。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对公开论文及所列官方来源；未知硬件和代码信息保留未知。研究结果为作者报告，未独立复现。", "timelineNote": "2016：在真实推物任务中展示动作条件视频预测可直接支持闭环控制。", "freshness": "基础奠基", "original": { "id": "arxiv-1610.00696", "originalSourceUrl": "https://arxiv.org/pdf/1610.00696v2", "versionedOriginalUrl": "https://arxiv.org/pdf/1610.00696v2", "finalDownloadUrl": "https://arxiv.org/pdf/1610.00696", "sha256": "c16bb0f44725a3a27307f790ffe31e032c21f4f90c9b182c0236266d3b783089", "bytes": 4135921, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:40.247254+00:00", "publicationDate": "2016/10/03", "publicationDateSourceUrl": "https://arxiv.org/abs/1610.00696", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1610.00696", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1610.00696", "pages": 8, "sourceIdentityStatus": "title_match", "sourceTitle": "Deep Visual Foresight for Planning Robot Motion", "sourceVersion": "1610.00696v2", "sourceVersionDate": "2017/03/13", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.506225+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1610.00696" }, "originalAnalysis": { "id": "arxiv-1610.00696", "sourceUrl": "https://arxiv.org/pdf/1610.00696", "sectionsRead": [ "§III–IV", "§V-A–C", "Table I", "Figure 6", "§VI" ], "methodsZh": "动作条件卷积LSTM预测图像变换核与合成掩模，无需物体模型或显式光流监督。把用户指定像素随预测流传播为位置分布，CEM选择使像素接近目标的动作；执行后用真实光流更新像素位置并重新规划。", "experimentsZh": "真实7自由度机械臂推动七件未见物体，比较随机动作、末端直接趋向目标及连续视觉伺服。使用64×64图像、三步约800毫秒预测时域，每约200毫秒重规划；训练数据来自十台具有不同相机布置的机器人。", "resultsZh": "作者报告最终目标像素距离2.52±1.06，对照连续伺服3.19±1.68、随机4.05±1.75；这是像素误差，不是成功率或毫米精度。双像素目标能诱导部分物体旋转，说明预测包含可用于控制的接触运动信息。", "limitationsZh": "任务集中于短时域反应式推物，候选时域内动作绑定不变，不能推断长期规划能力。机械臂遮挡导致像素跟踪漂到手臂，质量/接触预测偏差会把旋转变成平移；原文所读部分未明确厂商型号。", "reproductionZh": "需复现状态动作同步、视频预测训练、像素目标和光流跟踪，并用同一像素追踪协议比较基线。应同时检查遮挡失败、物体旋转与最终像素距离；本次未补猜硬件或运行代码。", "experimentType": "real", "robots": [ "7自由度机械臂（型号未明确）" ], "evidenceNotes": [ { "section/page": "§III–IV; Algorithm 1", "note": "动作条件像素流与CEM视觉MPC。" }, { "section/page": "§V-A–C; Table I; Figure 6", "note": "800毫秒时域、像素误差、遮挡与物理预测失败。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-1607.02565", "title": "Direct Sparse Odometry", "titleZh": "DSO：直接稀疏视觉里程计", "date": "2016-07-09", "datePrecision": "day", "year": 2016, "url": "https://arxiv.org/abs/1607.02565", "paperUrl": "https://arxiv.org/abs/1607.02565", "codeUrl": "https://github.com/JakobEngel/dso", "summary": "DSO以分布均匀的高梯度像素及小邻域光度残差联合优化相机位姿、逆深度、内参与亮度参数，显式建模曝光、响应曲线和暗角。采用滑动窗口与边缘化保持计算可控，不加稠密几何平滑先验，因此既直接又稀疏。", "abstractZh": "DSO以分布均匀的高梯度像素及小邻域光度残差联合优化相机位姿、逆深度、内参与亮度参数，显式建模曝光、响应曲线和暗角。采用滑动窗口与边缘化保持计算可控，不加稠密几何平滑先验，因此既直接又稀疏。\nTUM monoVO50段真实序列约105分钟、EuRoC MAV以及ICL-NUIM光线追踪合成序列；多次正向/反向运行，比较直接和特征方法、实时/低配置及标定消融，并人为加入光度/几何噪声。", "category": "视觉里程计 / 直接法", "tags": [ "DSO", "光度标定", "直接法", "滑动窗口", "真实数据评测" ], "directions": [ "导航与建图" ], "tier": "classic", "experimentType": "data", "robots": [], "robotFilters": [], "codeStatus": "open", "status": "实现已开源", "trainingNote": "无需策略训练；官方代码含数据集运行入口。应准备几何和光度标定，按 README 配置输入。", "contribution": "DSO以分布均匀的高梯度像素及小邻域光度残差联合优化相机位姿、逆深度、内参与亮度参数，显式建模曝光、响应曲线和暗角。采用滑动窗口与边缘化保持计算可控，不加稠密几何平滑先验，因此既直接又稀疏。", "whyUseful": "先准备曝光时间、响应/暗角以及准确几何标定，复现窗口/点数和实时约束；分别报告失跟踪比例、对齐后误差与未闭环漂移。不要只选良好纹理视频，应加入曝光变化、低视差和滚动快门压力测试。", "limitations": "这是单目视觉里程计，不包含完整全局回环和重定位；轨迹误差常经Sim(3)对齐，绝对尺度并未恢复。实时比较、可用标定和是否允许局部回环影响公平性；合成序列也不能替代机器人闭环导航。", "license": "GPL-3.0", "evidence": [ { "url": "https://arxiv.org/abs/1607.02565", "note": "首版日期与直接稀疏、光度校准模型的作者摘要。" }, { "url": "https://github.com/JakobEngel/dso", "note": "官方仓库核实 GPLv3、运行入口及无重定位、标定、快门和初始化限制。" }, { "url": "https://arxiv.org/pdf/1607.02565", "note": "2–3：光度模型、稀疏联合优化及滑窗。" }, { "url": "https://arxiv.org/pdf/1607.02565", "note": "4–5：三类数据、实时/噪声对照及几何标定边界。" }, { "url": "https://arxiv.org/pdf/1607.02565v2", "note": "2–3：光度模型、稀疏联合优化及滑窗。" }, { "url": "https://arxiv.org/pdf/1607.02565v2", "note": "4–5：三类数据、实时/噪声对照及几何标定边界。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "突出联合光度优化和相机响应建模，深化了直接法视觉里程计路线。", "experimentNote": "TUM monoVO50段真实序列约105分钟、EuRoC MAV以及ICL-NUIM光线追踪合成序列；多次正向/反向运行，比较直接和特征方法、实时/低配置及标定消融，并人为加入光度/几何噪声。", "projectUrl": null, "shortTitle": "", "original": { "id": "arxiv-1607.02565", "originalSourceUrl": "https://arxiv.org/pdf/1607.02565v2", "versionedOriginalUrl": "https://arxiv.org/pdf/1607.02565v2", "finalDownloadUrl": "https://arxiv.org/pdf/1607.02565", "sha256": "cd85260a38b0b5d0e82c85c37c3540c498a953e39a501d9af5192a35b8284223", "bytes": 7990289, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:45.251834+00:00", "publicationDate": "2016/07/09", "publicationDateSourceUrl": "https://arxiv.org/abs/1607.02565", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1607.02565", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1607.02565", "pages": 17, "sourceIdentityStatus": "title_match", "sourceTitle": "Direct Sparse Odometry", "sourceVersion": "1607.02565v2", "sourceVersionDate": "2016/10/07", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.514669+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1607.02565" }, "originalAnalysis": { "id": "arxiv-1607.02565", "sourceUrl": "https://arxiv.org/pdf/1607.02565v2", "sectionsRead": [ "2.1–2.3", "3", "4.1–4.3", "5", "Figs 10–12,23" ], "methodsZh": "DSO以分布均匀的高梯度像素及小邻域光度残差联合优化相机位姿、逆深度、内参与亮度参数，显式建模曝光、响应曲线和暗角。采用滑动窗口与边缘化保持计算可控，不加稠密几何平滑先验，因此既直接又稀疏。", "experimentsZh": "TUM monoVO50段真实序列约105分钟、EuRoC MAV以及ICL-NUIM光线追踪合成序列；多次正向/反向运行，比较直接和特征方法、实时/低配置及标定消融，并人为加入光度/几何噪声。", "resultsZh": "在标定良好数据上显示更强跟踪准确性和鲁棒性，低配置还能约五倍实时处理；完整光度标定有帮助，直接光度联合优化对部分亮度噪声更稳健。强几何噪声时结论相反，滚动快门和内参误差更伤害直接法。", "limitationsZh": "这是单目视觉里程计，不包含完整全局回环和重定位；轨迹误差常经Sim(3)对齐，绝对尺度并未恢复。实时比较、可用标定和是否允许局部回环影响公平性；合成序列也不能替代机器人闭环导航。", "reproductionZh": "先准备曝光时间、响应/暗角以及准确几何标定，复现窗口/点数和实时约束；分别报告失跟踪比例、对齐后误差与未闭环漂移。不要只选良好纹理视频，应加入曝光变化、低视差和滚动快门压力测试。", "experimentType": "data", "robots": [], "corrections": [ "包含真实和合成视觉数据评测，仍属data；不是已验证机器人闭环任务。" ], "evidenceNotes": [ { "note": "光度模型、稀疏联合优化及滑窗。", "section": "2–3" }, { "note": "三类数据、实时/噪声对照及几何标定边界。", "section": "4–5" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "1607.02565v2", "originalSha256": "cd85260a38b0b5d0e82c85c37c3540c498a953e39a501d9af5192a35b8284223" }, "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "cartographer-2016", "title": "Real-Time Loop Closure in 2D LIDAR SLAM", "titleZh": "Cartographer：二维激光 SLAM 中的实时回环", "date": "2016", "datePrecision": "year", "year": 2016, "url": "https://research.google/pubs/real-time-loop-closure-in-2d-lidar-slam/", "paperUrl": "https://research.google/pubs/real-time-loop-closure-in-2d-lidar-slam/", "codeUrl": "https://github.com/cartographer-project/cartographer", "summary": "Cartographer将连续激光扫描配准到小范围概率栅格子地图，局部用平滑占据概率的非线性优化；子地图完成后固定其栅格，预计算不同窗口内的最大值作为上界，以分支定界加速扫描到子地图回环搜索。后台优化扫描与子地…", "abstractZh": "Cartographer将连续激光扫描配准到小范围概率栅格子地图，局部用平滑占据概率的非线性优化；子地图完成后固定其栅格，预计算不同窗口内的最大值作为上界，以分支定界加速扫描到子地图回环搜索。后台优化扫描与子地图位姿图，纠正局部累计漂移。\n实验用德国博物馆背包采集数据、放在推车上的Neato吸尘器Revo LDS传感器及Radish公开序列。主要是录制传感器数据以在线算法重放；五条激光尺测距用于比较二维地图长度，公开序列用人工核验相对位姿关系评价，非自主机器人导航实验。", "category": "激光 SLAM / 子地图", "tags": [ "Cartographer", "二维 SLAM", "回环检测", "分支定界" ], "directions": [ "导航与建图" ], "tier": "classic", "experimentType": "real", "robots": [], "robotFilters": [], "codeStatus": "open", "status": "实现已开源", "trainingNote": "无需神经网络训练；Cartographer 源码和配置公开。官方 README 已声明不再积极维护，移植新 ROS 环境时需核对分支与依赖。", "contribution": "Cartographer将连续激光扫描配准到小范围概率栅格子地图，局部用平滑占据概率的非线性优化；子地图完成后固定其栅格，预计算不同窗口内的最大值作为上界，以分支定界加速扫描到子地图回环搜索。后台优化扫描与子地图位姿图，纠正局部累计漂移。", "whyUseful": "无需神经网络训练；复现须保留激光时间戳、IMU/里程计配置、搜索窗、子地图大小和优化权重。应分别报告CPU与墙钟、回环精确率及几何误差，不能把后续库的三维SLAM能力视作本篇验证。", "limitations": "二维投影依赖近似水平激光及重力估计，5厘米是栅格分辨率而非统一定位精度。各公开数据集分别调参，作者承认无法完全排除对场景过拟合；Freiburg hospital回环精确率77.3%，错误约束仍需鲁棒处理。", "license": "Apache-2.0", "evidence": [ { "url": "https://research.google/pubs/real-time-loop-closure-in-2d-lidar-slam/", "note": "Google 论文页确认 ICRA 2016、背包平台及二维实时回环。" }, { "url": "https://research.google.com/pubs/archive/45466.pdf", "note": "论文 PDF 支持子地图、分支定界与真实测试范围。" }, { "url": "https://github.com/cartographer-project/cartographer", "note": "官方仓库确认 Apache-2.0 和不再积极维护的声明。" }, { "url": "https://research.google.com/pubs/archive/45466.pdf", "note": "IV/V：Occupancy-grid submaps, nonlinear scan matching, precomputed max grids and scan-to-submap branch-and-bound." }, { "url": "https://research.google.com/pubs/archive/45466.pdf", "note": "VI-A PDF p5：1913s data,2253m estimated trajectory;360s wall clock;2.2GB;four loop-closure threads." }, { "url": "https://research.google.com/pubs/archive/45466.pdf", "note": "VI-B/C 表I/II/IV：Manually trolley-carried Revo LDS;five distance comparisons;dataset-specific tuning caveat;hospital loop precision77.3%." }, { "url": "https://research.google.com/pubs/archive/45466.pdf", "note": "PDF p5 图3/4：Visually inspected multiscale max grids and the Deutsches Museum floor map." } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "以子地图和高效全局匹配把回环计算纳入实时二维激光建图流程。", "experimentNote": "实验用德国博物馆背包采集数据、放在推车上的Neato吸尘器Revo LDS传感器及Radish公开序列。主要是录制传感器数据以在线算法重放；五条激光尺测距用于比较二维地图长度，公开序列用人工核验相对位姿关系评价，非自主机器人导航实验。", "projectUrl": null, "shortTitle": "", "original": { "id": "cartographer-2016", "originalSourceUrl": "https://research.google.com/pubs/archive/45466.pdf", "finalDownloadUrl": "https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45466.pdf", "sha256": "351edb45cd478c8cf2e22e375b8565a7abb3455e92f45e7a7f4163130bde69a2", "bytes": 1207803, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:34:55.290308+00:00", "publicationDate": "2016", "publicationDateSourceUrl": "https://research.google/pubs/real-time-loop-closure-in-2d-lidar-slam/", "license": "IEEE copyright; public author-hosted copy. Redistribution permission not verified.", "licenseSourceUrl": "https://research.google/pubs/real-time-loop-closure-in-2d-lidar-slam/", "licenseStatus": "rights_notice_verified", "licenseScope": "paper; does not establish code license", "pages": 8, "sourceIdentityStatus": "title_match_in_pdf", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.517180+00:00" }, "originalAnalysis": { "id": "cartographer-2016", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739401+00:00", "sourceUrl": "https://research.google.com/pubs/archive/45466.pdf", "sectionsRead": [ "IV 局部子地图和Ceres扫描匹配", "V 位姿图及分支定界回环", "VI-A–C 实测、Radish与表I–V", "PDF第5页图3/4可视检查" ], "methodsZh": "Cartographer将连续激光扫描配准到小范围概率栅格子地图，局部用平滑占据概率的非线性优化；子地图完成后固定其栅格，预计算不同窗口内的最大值作为上界，以分支定界加速扫描到子地图回环搜索。后台优化扫描与子地图位姿图，纠正局部累计漂移。", "experimentsZh": "实验用德国博物馆背包采集数据、放在推车上的Neato吸尘器Revo LDS传感器及Radish公开序列。主要是录制传感器数据以在线算法重放；五条激光尺测距用于比较二维地图长度，公开序列用人工核验相对位姿关系评价，非自主机器人导航实验。", "resultsZh": "作者报告1913秒、估计2253米的博物馆轨迹在Xeon E5-1650上360秒处理完，约5.3倍实时、峰值2.2GB；典型图优化约0.3秒。五条测距的相对误差绝对值0.2%至0.8%。Radish多数指标具竞争力，但MIT CSAIL不及所引Graph Mapping，回环精确率也非全为100%。", "limitationsZh": "二维投影依赖近似水平激光及重力估计，5厘米是栅格分辨率而非统一定位精度。各公开数据集分别调参，作者承认无法完全排除对场景过拟合；Freiburg hospital回环精确率77.3%，错误约束仍需鲁棒处理。", "reproductionZh": "无需神经网络训练；复现须保留激光时间戳、IMU/里程计配置、搜索窗、子地图大小和优化权重。应分别报告CPU与墙钟、回环精确率及几何误差，不能把后续库的三维SLAM能力视作本篇验证。", "evidenceNotes": [ { "section/page": "IV/V", "note": "Occupancy-grid submaps, nonlinear scan matching, precomputed max grids and scan-to-submap branch-and-bound." }, { "section/page": "VI-A PDF p5", "note": "1913s data,2253m estimated trajectory;360s wall clock;2.2GB;four loop-closure threads." }, { "section/page": "VI-B/C 表I/II/IV", "note": "Manually trolley-carried Revo LDS;five distance comparisons;dataset-specific tuning caveat;hospital loop precision77.3%." }, { "section/page": "PDF p5 图3/4", "note": "Visually inspected multiscale max grids and the Deutsches Museum floor map." } ], "experimentType": "real", "robots": [], "corrections": [ "5厘米指二维栅格分辨率，不是所有场景定位误差。", "使用真实传感器采集与数据重放；没有证实自主机器人闭环导航。" ], "sourceVersion": null, "originalSha256": "351edb45cd478c8cf2e22e375b8565a7abb3455e92f45e7a7f4163130bde69a2" }, "analysisVerifiedAt": "2026-10-04T13:47:58.739401+00:00" }, { "id": "arxiv-1509.02971", "title": "Continuous control with deep reinforcement learning", "titleZh": "利用深度强化学习进行连续控制", "shortTitle": "DDPG", "date": "2015-09-09", "datePrecision": "day", "year": 2015, "url": "https://arxiv.org/abs/1509.02971", "codeUrl": null, "category": "强化学习", "tags": [ "连续动作", "Actor-Critic", "基础算法" ], "tier": "foundation", "summary": "DDPG将确定性策略梯度与深度actor-critic结合，actor输出连续动作，critic拟合Bellman目标。经验回放降低时序相关，缓慢软更新的双目标网络稳定训练；低维输入使用batch normal…", "abstractZh": "DDPG将确定性策略梯度与深度actor-critic结合，actor输出连续动作，critic拟合Bellman目标。经验回放降低时序相关，缓慢软更新的双目标网络稳定训练；低维输入使用batch normalization，探索加入Ornstein-Uhlenbeck相关噪声。\n在多种MuJoCo连续控制及TORCS任务测试状态和像素输入，像素为连续三帧64×64RGB、动作重复三次。每环境五次独立运行，最多250万步，周期性无探索噪声评估；以随机策略0和具真实动力学的iLQG规划器1归一化。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "sim", "robots": [], "codeStatus": "unknown", "status": "本次未核实原作者原始实验代码。", "trainingNote": "仿真在线采样，回放池复用历史数据；行为策略加入探索噪声。", "whyUseful": "复现需固定原始环境、奖励、动作重复、探索噪声与目标更新率，不能把现代库默认DDPG当原论文配置。比较iLQG时要注明其已知模型优势；本轮只核读原论文，未核验原作者完整训练代码或重新跑种子。", "contribution": "DDPG将确定性策略梯度与深度actor-critic结合，actor输出连续动作，critic拟合Bellman目标。经验回放降低时序相关，缓慢软更新的双目标网络稳定训练；低维输入使用batch normalization，探索加入Ornstein-Uhlenbeck相关噪声。", "limitations": "分数是相对回报，不是成功率；最好种子与五次均值应区分。非线性函数近似没有收敛保证，仍需大量试错回合。全为仿真，论文提及机械臂/人形任务不表示真实平台安全部署。", "caveats": "仿真人形、机械臂等任务不代表使用了同名实体机器人。", "evidence": [ { "url": "https://arxiv.org/abs/1509.02971", "note": "首发日期；摘要明确二十多个simulated physics tasks。" }, { "url": "https://arxiv.org/pdf/1509.02971", "note": "算法结构、经验回放、目标网络与探索设置。" }, { "url": "https://arxiv.org/pdf/1509.02971", "note": "§4/Table 1：五次运行，平均/最好分开；随机0/iLQG1归一化。" }, { "url": "https://arxiv.org/pdf/1509.02971", "note": "§3：软目标、回放、归一化、OU探索。" }, { "url": "https://arxiv.org/pdf/1509.02971", "note": "§6：无收敛保证且样本需求大。" } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "基础理论", "强化学习" ], "robotFilters": [], "verificationNote": "2026-10-04核对原论文元数据与实验范围；不将第三方复现标为作者原版。", "original": { "id": "arxiv-1509.02971", "originalSourceUrl": "https://arxiv.org/pdf/1509.02971v6", "versionedOriginalUrl": "https://arxiv.org/pdf/1509.02971v6", "finalDownloadUrl": "https://arxiv.org/pdf/1509.02971", "sha256": "34ef677ecdee8c2a72721c70c7d849191be72af39ac8691a12e8c2979a28feda", "bytes": 663643, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:45.150218+00:00", "publicationDate": "2015/09/09", "publicationDateSourceUrl": "https://arxiv.org/abs/1509.02971", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1509.02971", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1509.02971", "pages": 14, "sourceIdentityStatus": "title_match", "sourceTitle": "Continuous control with deep reinforcement learning", "sourceVersion": "1509.02971v6", "sourceVersionDate": "2019/07/05", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.518977+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1509.02971" }, "originalAnalysis": { "id": "arxiv-1509.02971", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090155+00:00", "sourceUrl": "https://arxiv.org/pdf/1509.02971", "sectionsRead": [ "§3", "§4", "Table 1", "§6", "Supplementary architecture" ], "methodsZh": "DDPG将确定性策略梯度与深度actor-critic结合，actor输出连续动作，critic拟合Bellman目标。经验回放降低时序相关，缓慢软更新的双目标网络稳定训练；低维输入使用batch normalization，探索加入Ornstein-Uhlenbeck相关噪声。", "experimentsZh": "在多种MuJoCo连续控制及TORCS任务测试状态和像素输入，像素为连续三帧64×64RGB、动作重复三次。每环境五次独立运行，最多250万步，周期性无探索噪声评估；以随机策略0和具真实动力学的iLQG规划器1归一化。", "resultsZh": "作者表1显示不少状态任务接近或超过规划参照，例如cheetah平均0.903、hardCheetah1.311；相应像素为0.457/1.204。不是所有像素任务都成功，某些均值低于随机；消融支持目标网络和归一化的重要性。", "limitationsZh": "分数是相对回报，不是成功率；最好种子与五次均值应区分。非线性函数近似没有收敛保证，仍需大量试错回合。全为仿真，论文提及机械臂/人形任务不表示真实平台安全部署。", "reproductionZh": "复现需固定原始环境、奖励、动作重复、探索噪声与目标更新率，不能把现代库默认DDPG当原论文配置。比较iLQG时要注明其已知模型优势；本轮只核读原论文，未核验原作者完整训练代码或重新跑种子。", "experimentType": "sim", "robots": [], "evidenceNotes": [ { "section/page": "§4/Table 1", "note": "五次运行，平均/最好分开；随机0/iLQG1归一化。" }, { "section/page": "§3", "note": "软目标、回放、归一化、OU探索。" }, { "section/page": "§6", "note": "无收敛保证且样本需求大。" } ] }, "experimentNote": "在多种MuJoCo连续控制及TORCS任务测试状态和像素输入，像素为连续三帧64×64RGB、动作重复三次。每环境五次独立运行，最多250万步，周期性无探索噪声评估；以随机策略0和具真实动力学的iLQG规划器1归一化。", "analysisVerifiedAt": "2026-10-04T13:50:35.090155+00:00" }, { "id": "arxiv-1504.00702", "title": "End-to-End Training of Deep Visuomotor Policies", "titleZh": "深度视觉运动策略的端到端训练", "shortTitle": "Deep Visuomotor Policies", "date": "2015-04-02", "datePrecision": "day", "year": 2015, "url": "https://arxiv.org/abs/1504.00702", "codeUrl": "https://github.com/cbfinn/gps", "category": "视觉控制", "tags": [ "GPS", "端到端", "操作" ], "tier": "classic", "summary": "引导式策略搜索把局部轨迹优化与视觉策略监督学习交替进行，训练时利用完整状态，执行时只用相机与本体。CNN空间softmax把特征图转为可微坐标点，再映射到控制量；先预训练视觉和局部控制器，最后联合端到端优化。", "abstractZh": "引导式策略搜索把局部轨迹优化与视觉策略监督学习交替进行，训练时利用完整状态，执行时只用相机与本体。CNN空间softmax把特征图转为可微坐标点，再映射到控制量；先预训练视觉和局部控制器，最后联合端到端优化。\n除仿真策略搜索比较外，在PR2测试挂衣架、形状盒插块、锤爪配准和拧瓶盖。评测训练位置、新位置/握姿及视觉杂物，目标位置变化约10–20厘米；各任务使用156–288次五秒控制试验，另有约1000张视觉预训练图。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "PR2" ], "codeStatus": "open", "status": "作者团队GPS重实现公开；历史依赖较旧。", "trainingNote": "训练阶段需要比执行阶段更丰富的状态信息与轨迹控制器。", "whyUseful": "保留空间softmax、状态可见性划分、局部控制器约束和视觉预训练，逐项报告试验数与采集成本。本次视觉核对PDF图9后引用成绩，未执行历史训练栈。", "contribution": "引导式策略搜索把局部轨迹优化与视觉策略监督学习交替进行，训练时利用完整状态，执行时只用相机与本体。CNN空间softmax把特征图转为可微坐标点，再映射到控制量；先预训练视觉和局部控制器，最后联合端到端优化。", "limitations": "训练需要仪器化完整状态及受限任务分布，大幅背景改变或遮挡会失败；不能从端到端推断无预训练或无状态监督。PR2实际为effort接口而非理想闭环力矩控制，硬件接口需区别。", "caveats": "2015为预印本，JMLR发表于2016。PR2实际提供effort接口，并非理想闭环力矩控制。", "evidence": [ { "url": "https://arxiv.org/abs/1504.00702", "note": "首发日期与方法摘要。" }, { "url": "https://arxiv.org/pdf/1504.00702", "note": "第6节PR2实验；附录脚注5说明effort接口。" }, { "url": "https://rll.berkeley.edu/gps/", "note": "作者团队文档说明重实现、PR2/ROS接口和GitHub代码。" }, { "url": "https://arxiv.org/pdf/1504.00702", "note": "§4–5; Figure 2：完整状态引导、空间softmax与端到端策略。" }, { "url": "https://arxiv.org/pdf/1504.00702", "note": "PDF p24 Figure 9; Table 4; footnote 5：可视核对新位置/干扰成功率、156–288试验及effort接口。" } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "操作与抓取" ], "robotFilters": [ "PR2" ], "verificationNote": "2026-10-04核对论文与作者实现文档；代码未运行。", "original": { "id": "arxiv-1504.00702", "originalSourceUrl": "https://arxiv.org/pdf/1504.00702v5", "versionedOriginalUrl": "https://arxiv.org/pdf/1504.00702v5", "finalDownloadUrl": "https://arxiv.org/pdf/1504.00702", "sha256": "28ad13a058680cbf5c007c7fdf2f66330bef4b3efa7d1c06169680bc81758f41", "bytes": 6257924, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:53.409517+00:00", "publicationDate": "2015/04/02", "publicationDateSourceUrl": "https://arxiv.org/abs/1504.00702", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1504.00702", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1504.00702", "pages": 40, "sourceIdentityStatus": "title_match", "sourceTitle": "End-to-End Training of Deep Visuomotor Policies", "sourceVersion": "1504.00702v5", "sourceVersionDate": "2016/04/19", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.525774+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1504.00702" }, "originalAnalysis": { "id": "arxiv-1504.00702", "sourceUrl": "https://arxiv.org/pdf/1504.00702", "sectionsRead": [ "§4–5", "§6.1,6.4–6.6", "Figure 9", "Table 4", "§7", "Appendix B footnote 5" ], "methodsZh": "引导式策略搜索把局部轨迹优化与视觉策略监督学习交替进行，训练时利用完整状态，执行时只用相机与本体。CNN空间softmax把特征图转为可微坐标点，再映射到控制量；先预训练视觉和局部控制器，最后联合端到端优化。", "experimentsZh": "除仿真策略搜索比较外，在PR2测试挂衣架、形状盒插块、锤爪配准和拧瓶盖。评测训练位置、新位置/握姿及视觉杂物，目标位置变化约10–20厘米；各任务使用156–288次五秒控制试验，另有约1000张视觉预训练图。", "resultsZh": "作者报告端到端方式优于冻结位置特征或显式位姿预测。图9中插块新位置成功91.7%、视觉干扰87.5%；瓶盖对应83.3%、62.5%，显示视觉变化仍明显削弱效果。约15分钟实机试验不含全部图像采集和3–4小时计算训练。", "limitationsZh": "训练需要仪器化完整状态及受限任务分布，大幅背景改变或遮挡会失败；不能从端到端推断无预训练或无状态监督。PR2实际为effort接口而非理想闭环力矩控制，硬件接口需区别。", "reproductionZh": "保留空间softmax、状态可见性划分、局部控制器约束和视觉预训练，逐项报告试验数与采集成本。本次视觉核对PDF图9后引用成绩，未执行历史训练栈。", "experimentType": "both", "robots": [ "PR2" ], "evidenceNotes": [ { "section/page": "§4–5; Figure 2", "note": "完整状态引导、空间softmax与端到端策略。" }, { "section/page": "PDF p24 Figure 9; Table 4; footnote 5", "note": "可视核对新位置/干扰成功率、156–288试验及effort接口。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "除仿真策略搜索比较外，在PR2测试挂衣架、形状盒插块、锤爪配准和拧瓶盖。评测训练位置、新位置/握姿及视觉杂物，目标位置变化约10–20厘米；各任务使用156–288次五秒控制试验，另有约1000张视觉预训练图。", "analysisVerifiedAt": "2026-10-04" }, { "id": "arxiv-1502.00956", "title": "ORB-SLAM: a Versatile and Accurate Monocular SLAM System", "titleZh": "ORB-SLAM：通用且精确的单目同步定位与建图系统", "date": "2015-02-03", "datePrecision": "day", "year": 2015, "url": "https://arxiv.org/abs/1502.00956", "paperUrl": "https://arxiv.org/abs/1502.00956", "codeUrl": "https://github.com/raulmur/ORB_SLAM", "summary": "ORB-SLAM用同一ORB特征贯穿跟踪、局部地图、词袋重定位和回环；三个线程协作，关键帧/地图点剔除控制冗余。初始化比较平面单应和一般场景基本矩阵，局部BA优化几何，回环以Sim(3)及Essential G…", "abstractZh": "ORB-SLAM用同一ORB特征贯穿跟踪、局部地图、词袋重定位和回环；三个线程协作，关键帧/地图点剔除控制冗余。初始化比较平面单应和一般场景基本矩阵，局部BA优化几何，回环以Sim(3)及Essential Graph修正尺度漂移。\n在NewCollege机器人视频、TUM RGB-D的16条室内手持序列及KITTI10条车载序列验证共27序列，但算法只使用单目图像。评估定位、重定位、长期重复访问地图增长、实时性及回环优化方式。", "category": "视觉 SLAM / 特征法", "tags": [ "ORB", "单目 SLAM", "回环检测", "关键帧", "真实数据评测" ], "directions": [ "导航与建图" ], "tier": "classic", "experimentType": "data", "robots": [], "robotFilters": [], "codeStatus": "open", "status": "实现已开源", "trainingNote": "无需神经网络训练；官方 C++/ROS 实现及依赖说明公开。需相机标定、特征词袋与兼容构建环境。", "contribution": "ORB-SLAM用同一ORB特征贯穿跟踪、局部地图、词袋重定位和回环；三个线程协作，关键帧/地图点剔除控制冗余。初始化比较平面单应和一般场景基本矩阵，局部BA优化几何，回环以Sim(3)及Essential Graph修正尺度漂移。", "whyUseful": "复现相机内参、ORB词袋、线程与关键帧阈值，并分别测无回环里程计和完整SLAM。先跑同一单目输入与尺度对齐评测，再测试失跟踪恢复；不能引用后续ORB-SLAM2/3的双目、惯性能力到本篇。", "limitations": "单目绝对尺度不确定，初始化需要足够视差；稀疏角点地图不等于稠密可通行地图，模糊与弱纹理会影响特征。记录视频成功不能等同自主导航安全，文中精度须保留对齐条件和失败序列背景。", "license": "GPL-3.0", "evidence": [ { "url": "https://arxiv.org/abs/1502.00956", "note": "arXiv 核对首版日期、系统能力及 27 段数据集评估。" }, { "url": "https://github.com/raulmur/ORB_SLAM", "note": "作者仓库确认公开实现、GPLv3、数据和构建说明。" }, { "url": "https://arxiv.org/pdf/1502.00956", "note": "III–VII：统一ORB、三线程、初始化及Sim(3)回环。" }, { "url": "https://arxiv.org/pdf/1502.00956", "note": "VIII–IX：27序列、尺度对齐精度及地图/关键帧能力边界。" }, { "url": "https://arxiv.org/pdf/1502.00956v2", "note": "III–VII：统一ORB、三线程、初始化及Sim(3)回环。" }, { "url": "https://arxiv.org/pdf/1502.00956v2", "note": "VIII–IX：27序列、尺度对齐精度及地图/关键帧能力边界。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "以统一特征和完整回环/重定位流程，成为特征型视觉 SLAM 的代表性基线。", "experimentNote": "在NewCollege机器人视频、TUM RGB-D的16条室内手持序列及KITTI10条车载序列验证共27序列，但算法只使用单目图像。评估定位、重定位、长期重复访问地图增长、实时性及回环优化方式。", "projectUrl": null, "shortTitle": "", "original": { "id": "arxiv-1502.00956", "originalSourceUrl": "https://arxiv.org/pdf/1502.00956v2", "versionedOriginalUrl": "https://arxiv.org/pdf/1502.00956v2", "finalDownloadUrl": "https://arxiv.org/pdf/1502.00956", "sha256": "442bcef21b04c007628c452e21b3122296e76fee727b1fc766e7d2b666ec3634", "bytes": 4202706, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:57.039400+00:00", "publicationDate": "2015/02/03", "publicationDateSourceUrl": "https://arxiv.org/abs/1502.00956", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1502.00956", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1502.00956", "pages": 18, "sourceIdentityStatus": "title_match", "sourceTitle": "ORB-SLAM: a Versatile and Accurate Monocular SLAM System", "sourceVersion": "1502.00956v2", "sourceVersionDate": "2015/09/18", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.531881+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1502.00956" }, "originalAnalysis": { "id": "arxiv-1502.00956", "sourceUrl": "https://arxiv.org/pdf/1502.00956v2", "sectionsRead": [ "III", "IV", "V–VII", "VIII-A–F", "IX", "Tables I–VI" ], "methodsZh": "ORB-SLAM用同一ORB特征贯穿跟踪、局部地图、词袋重定位和回环；三个线程协作，关键帧/地图点剔除控制冗余。初始化比较平面单应和一般场景基本矩阵，局部BA优化几何，回环以Sim(3)及Essential Graph修正尺度漂移。", "experimentsZh": "在NewCollege机器人视频、TUM RGB-D的16条室内手持序列及KITTI10条车载序列验证共27序列，但算法只使用单目图像。评估定位、重定位、长期重复访问地图增长、实时性及回环优化方式。", "resultsZh": "作者报告经尺度对齐后小室内通常亚厘米、大室外数米级；重定位可处理较大视角变化，回访剔除冗余关键帧避免地图持续膨胀。回环先做稀疏图优化再BA比直接从大漂移状态做全BA更容易收敛。", "limitationsZh": "单目绝对尺度不确定，初始化需要足够视差；稀疏角点地图不等于稠密可通行地图，模糊与弱纹理会影响特征。记录视频成功不能等同自主导航安全，文中精度须保留对齐条件和失败序列背景。", "reproductionZh": "复现相机内参、ORB词袋、线程与关键帧阈值，并分别测无回环里程计和完整SLAM。先跑同一单目输入与尺度对齐评测，再测试失跟踪恢复；不能引用后续ORB-SLAM2/3的双目、惯性能力到本篇。", "experimentType": "data", "robots": [], "corrections": [ "使用TUM RGB-D数据集时本方法仍是单目；不能把RGB-D/惯性能力从后续版本移入原文。" ], "evidenceNotes": [ { "note": "统一ORB、三线程、初始化及Sim(3)回环。", "section": "III–VII" }, { "note": "27序列、尺度对齐精度及地图/关键帧能力边界。", "section": "VIII–IX" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "1502.00956v2", "originalSha256": "442bcef21b04c007628c452e21b3122296e76fee727b1fc766e7d2b666ec3634" }, "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "loam-2014", "title": "LOAM: Lidar Odometry and Mapping in Real-time", "titleZh": "LOAM：实时激光雷达里程计与建图", "date": "2014", "datePrecision": "year", "year": 2014, "url": "https://www.roboticsproceedings.org/rss10/p07.html", "paperUrl": "https://www.roboticsproceedings.org/rss10/p07.html", "codeUrl": null, "summary": "LOAM把点云运动估计拆为约10Hz里程计和约1Hz精细建图，利用扫描邻域曲率选边缘/平面点，通过点线与点面距离估计相邻扫描运动并去畸变，再与地图匹配；融合两个位姿变换兼顾速度和低漂移。", "abstractZh": "LOAM把点云运动估计拆为约10Hz里程计和约1Hz精细建图，利用扫描邻域曲率选边缘/平面点，通过点线与点面距离估计相邻扫描运动并去畸变，再与地图匹配；融合两个位姿变换兼顾速度和低漂移。\n旋转Hokuyo在室内推车、室外地面车辆上以0.5米/秒采集走廊、大厅、植被道路和果园；闭环端点或GPS/INS作参照。另比较Xsens MTi-10辅助的手持运动，KITTI分支仅用Velodyne雷达数据。", "category": "激光 SLAM / 里程计", "tags": [ "LiDAR", "点云配准", "运动畸变", "LOAM" ], "directions": [ "导航与建图" ], "tier": "foundation", "experimentType": "real", "robots": [ "推车与地面车辆平台（型号未给出）" ], "robotFilters": [], "codeStatus": "unknown", "status": "官方代码待核实", "trainingNote": "几何系统，无策略训练；本轮未核实仍可访问的作者原始源码与许可证。未把社区改写版冒充论文官方实现。", "contribution": "LOAM把点云运动估计拆为约10Hz里程计和约1Hz精细建图，利用扫描邻域曲率选边缘/平面点，通过点线与点面距离估计相邻扫描运动并去畸变，再与地图匹配；融合两个位姿变换兼顾速度和低漂移。", "whyUseful": "复现应分别设置Hokuyo和KITTI扫描模式、点特征阈值、时间插值及IMU预处理，不能把后来LOAM衍生物的回环加入原论文声明。这里无神经训练；未核定原始发布代码与依赖，且不从传感器推断车辆厂牌。", "limitations": "低漂移不是零漂移，原始系统尚未包含全局回环纠正，作者将其列为后续工作。结果依赖扫描结构和足够稳定几何，缓慢地面采集不能代表激烈运动；部分参考真值来自卷尺，精度口径有限。", "license": "未核实（原始作者实现）", "evidence": [ { "url": "https://www.roboticsproceedings.org/rss10/p07.html", "note": "RSS 官方论文页确认题名、2014 年、双频率架构和 KITTI/真实实验范围。" }, { "url": "https://publications.ri.cmu.edu/loam-lidar-odometry-and-mapping-in-real-time", "note": "作者机构论文页独立核对发表年与方法摘要。" }, { "url": "https://www.roboticsproceedings.org/rss10/p07.pdf", "note": "Fig.3：10Hz里程计/1Hz建图，输出位姿融合。" }, { "url": "https://www.roboticsproceedings.org/rss10/p07.pdf", "note": "Tables I–II/§VII-C：真实漂移与0.88% KITTI报告，非现代榜单。" }, { "url": "https://www.roboticsproceedings.org/rss10/p07.pdf", "note": "§VIII：回环修正列为未来工作。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "以高频里程计、低频精细建图的分工建立了激光 SLAM 的重要工程范式。", "experimentNote": "旋转Hokuyo在室内推车、室外地面车辆上以0.5米/秒采集走廊、大厅、植被道路和果园；闭环端点或GPS/INS作参照。另比较Xsens MTi-10辅助的手持运动，KITTI分支仅用Velodyne雷达数据。", "projectUrl": null, "shortTitle": "", "original": { "id": "loam-2014", "originalSourceUrl": "https://www.roboticsproceedings.org/rss10/p07.pdf", "finalDownloadUrl": "https://www.roboticsproceedings.org/rss10/p07.pdf", "sha256": "0d66f17eadd7770e5894df92a3ca854a9a53f366f2aec6a828eb1bbcd6e28941", "bytes": 3645860, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:12.791496+00:00", "publicationDate": "2014-07", "publicationDateSourceUrl": "https://www.roboticsproceedings.org/rss10/p07.html", "license": null, "licenseSourceUrl": null, "licenseStatus": "not_verified", "pages": 9, "sourceIdentityStatus": "title_match_in_pdf", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.545350+00:00" }, "originalAnalysis": { "id": "loam-2014", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090156+00:00", "sourceUrl": "https://www.roboticsproceedings.org/rss10/p07.pdf", "sectionsRead": [ "§IV-B", "§V–VI", "§VII-A–C", "Tables I–II", "§VIII" ], "methodsZh": "LOAM把点云运动估计拆为约10Hz里程计和约1Hz精细建图，利用扫描邻域曲率选边缘/平面点，通过点线与点面距离估计相邻扫描运动并去畸变，再与地图匹配；融合两个位姿变换兼顾速度和低漂移。", "experimentsZh": "旋转Hokuyo在室内推车、室外地面车辆上以0.5米/秒采集走廊、大厅、植被道路和果园；闭环端点或GPS/INS作参照。另比较Xsens MTi-10辅助的手持运动，KITTI分支仅用Velodyne雷达数据。", "resultsZh": "作者表I走廊漂移0.9–1.1%、果园2.3–2.8%；表II IMU预处理可进一步改善非线性运动。KITTI当时提交总体位置误差0.88%，按100–800米轨迹段统计，排名是发表时状态而非当前榜单。", "limitationsZh": "低漂移不是零漂移，原始系统尚未包含全局回环纠正，作者将其列为后续工作。结果依赖扫描结构和足够稳定几何，缓慢地面采集不能代表激烈运动；部分参考真值来自卷尺，精度口径有限。", "reproductionZh": "复现应分别设置Hokuyo和KITTI扫描模式、点特征阈值、时间插值及IMU预处理，不能把后来LOAM衍生物的回环加入原论文声明。这里无神经训练；未核定原始发布代码与依赖，且不从传感器推断车辆厂牌。", "experimentType": "real", "robots": [ "推车与地面车辆平台（型号未给出）" ], "evidenceNotes": [ { "section/page": "Fig.3", "note": "10Hz里程计/1Hz建图，输出位姿融合。" }, { "section/page": "Tables I–II/§VII-C", "note": "真实漂移与0.88% KITTI报告，非现代榜单。" }, { "section/page": "§VIII", "note": "回环修正列为未来工作。" } ] }, "analysisVerifiedAt": "2026-10-04T13:50:35.090156+00:00" }, { "id": "lsd-slam-2014", "title": "LSD-SLAM: Large-Scale Direct Monocular SLAM", "titleZh": "LSD-SLAM：大规模直接法单目同步定位与建图", "date": "2014", "datePrecision": "year", "year": 2014, "url": "https://jakobengel.github.io/pdf/engel14eccv.pdf", "paperUrl": "https://jakobengel.github.io/pdf/engel14eccv.pdf", "codeUrl": "https://github.com/tum-vision/lsd_slam", "summary": "LSD-SLAM用有梯度像素构建半稠密逆深度及不确定度，不先提取稀疏特征。新帧通过方差归一化光度误差估计SE(3)位姿，多次小基线立体匹配更新深度；关键帧间联合光度与深度残差估计Sim(3)，在位姿图中显式修正…", "abstractZh": "LSD-SLAM用有梯度像素构建半稠密逆深度及不确定度，不先提取稀疏特征。新帧通过方差归一化光度误差估计SE(3)位姿，多次小基线立体匹配更新深度；关键帧间联合光度与深度残差估计Sim(3)，在位姿图中显式修正尺度漂移和回环。\n作者在约500米、六分钟手持单目户外轨迹展示大尺度回环，并以TUM RGB-D的RGB序列和两条合成序列进行定量比较；另测金字塔层数与ESM对对齐收敛范围的影响。TUM评测实际使用首帧深度初始化和设定初始尺度，之后采用单目跟踪。", "category": "视觉 SLAM / 直接法", "tags": [ "直接法", "半稠密建图", "单目 SLAM", "尺度漂移", "真实数据评测" ], "directions": [ "导航与建图" ], "tier": "classic", "experimentType": "data", "robots": [], "robotFilters": [], "codeStatus": "open", "status": "实现已开源", "trainingNote": "无需神经网络训练；官方 ROS 实现公开，README 给出历史 Ubuntu/ROS 环境、输入序列及相机标定要求。", "contribution": "LSD-SLAM用有梯度像素构建半稠密逆深度及不确定度，不先提取稀疏特征。新帧通过方差归一化光度误差估计SE(3)位姿，多次小基线立体匹配更新深度；关键帧间联合光度与深度残差估计Sim(3)，在位姿图中显式修正尺度漂移和回环。", "whyUseful": "无需训练，重点是相机标定、图像金字塔、深度方差传播、Huber核及Sim(3)图优化。复现应区分纯单目随机初始化和用首帧深度的基准设置，固定尺度对齐规则，并记录失败序列，避免仅比较成功运行时的轨迹误差。", "limitations": "单目本身不能观测绝对尺度；直接对齐是非凸问题，仍需良好初始化、足够平移和稳定成像。小规模选定序列不能证明所有场景鲁棒，随机深度自初始化的系统评估被明确留给未来；本文无机器人自主闭环实验。", "license": "GPL-3.0", "evidence": [ { "url": "https://jakobengel.github.io/pdf/engel14eccv.pdf", "note": "作者论文 PDF 核对 2014 年方法与半稠密直接建图范围。" }, { "url": "https://cvg.cit.tum.de/research/vslam/lsdslam", "note": "作者实验室页提供论文、真实输入序列与代码入口。" }, { "url": "https://github.com/tum-vision/lsd_slam", "note": "官方实现确认 GPLv3、ROS 依赖、Sim3 输出和标定/运动注意事项。" }, { "url": "https://jakobengel.github.io/pdf/engel14eccv.pdf", "note": "3.1/3.3/3.5：Random-depth initialization possible;uncertainty-weighted photometric tracking;depth residual required to constrain relative scale." }, { "url": "https://jakobengel.github.io/pdf/engel14eccv.pdf", "note": "4.2 PDF p12：First depth map used to bootstrap and obtain initial scale on RGB-D benchmark." }, { "url": "https://jakobengel.github.io/pdf/engel14eccv.pdf", "note": "PDF p13 图8/9：Visual inspection confirms before/after scale alignment and RMSE units in cm:4.52,1.47,0.04,0.35." }, { "url": "https://jakobengel.github.io/pdf/engel14eccv.pdf", "note": "4.3：ESM/more pyramid levels enlarge convergence radius, not notably final accuracy." } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "把半稠密直接法从局部里程计推进到带尺度校正和回环的大范围地图。", "experimentNote": "作者在约500米、六分钟手持单目户外轨迹展示大尺度回环，并以TUM RGB-D的RGB序列和两条合成序列进行定量比较；另测金字塔层数与ESM对对齐收敛范围的影响。TUM评测实际使用首帧深度初始化和设定初始尺度，之后采用单目跟踪。", "projectUrl": null, "shortTitle": "", "original": { "id": "lsd-slam-2014", "originalSourceUrl": "https://jakobengel.github.io/pdf/engel14eccv.pdf", "finalDownloadUrl": "https://jakobengel.github.io/pdf/engel14eccv.pdf", "sha256": "9d39989b15d1847d51432e90ba02cf5c7ae2b04da478214cf3fa050d2c39c64a", "bytes": 11044034, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:03.754674+00:00", "publicationDate": null, "publicationDateSourceUrl": null, "license": null, "licenseSourceUrl": null, "licenseStatus": "not_verified", "pages": 16, "sourceIdentityStatus": "title_match_in_pdf", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.541917+00:00" }, "originalAnalysis": { "id": "lsd-slam-2014", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739403+00:00", "sourceUrl": "https://jakobengel.github.io/pdf/engel14eccv.pdf", "sectionsRead": [ "3.1–3.6 完整方法、深度/位姿图和尺度感知对齐", "4.1–4.3 长轨迹、TUM与收敛域", "5 结论", "PDF第13页图8/9可视检查" ], "methodsZh": "LSD-SLAM用有梯度像素构建半稠密逆深度及不确定度，不先提取稀疏特征。新帧通过方差归一化光度误差估计SE(3)位姿，多次小基线立体匹配更新深度；关键帧间联合光度与深度残差估计Sim(3)，在位姿图中显式修正尺度漂移和回环。", "experimentsZh": "作者在约500米、六分钟手持单目户外轨迹展示大尺度回环，并以TUM RGB-D的RGB序列和两条合成序列进行定量比较；另测金字塔层数与ESM对对齐收敛范围的影响。TUM评测实际使用首帧深度初始化和设定初始尺度，之后采用单目跟踪。", "resultsZh": "图9报告绝对轨迹RMSE：fr2/desk为4.52厘米、fr2/xyz为1.47厘米；合成desk为0.04厘米。相对所比半稠密里程计有改善，但部分RGB-D方法误差更小；图8中回环消除不同尺度重复结构，说明地图一致性而非恢复绝对尺度。", "limitationsZh": "单目本身不能观测绝对尺度；直接对齐是非凸问题，仍需良好初始化、足够平移和稳定成像。小规模选定序列不能证明所有场景鲁棒，随机深度自初始化的系统评估被明确留给未来；本文无机器人自主闭环实验。", "reproductionZh": "无需训练，重点是相机标定、图像金字塔、深度方差传播、Huber核及Sim(3)图优化。复现应区分纯单目随机初始化和用首帧深度的基准设置，固定尺度对齐规则，并记录失败序列，避免仅比较成功运行时的轨迹误差。", "evidenceNotes": [ { "section/page": "3.1/3.3/3.5", "note": "Random-depth initialization possible;uncertainty-weighted photometric tracking;depth residual required to constrain relative scale." }, { "section/page": "4.2 PDF p12", "note": "First depth map used to bootstrap and obtain initial scale on RGB-D benchmark." }, { "section/page": "PDF p13 图8/9", "note": "Visual inspection confirms before/after scale alignment and RMSE units in cm:4.52,1.47,0.04,0.35." }, { "section/page": "4.3", "note": "ESM/more pyramid levels enlarge convergence radius, not notably final accuracy." } ], "experimentType": "data", "robots": [], "corrections": [ "定量基准使用首帧深度初始化；不应称该比较全程从零使用纯RGB。", "实验包括真实与合成图像序列，data表示离线视觉SLAM评测而非实机任务。" ], "sourceVersion": null, "originalSha256": "9d39989b15d1847d51432e90ba02cf5c7ae2b04da478214cf3fa050d2c39c64a" }, "analysisVerifiedAt": "2026-10-04T13:47:58.739403+00:00" }, { "id": "dmp", "title": "Dynamical Movement Primitives: Learning Attractor Models for Motor Behaviors", "titleZh": "动力学运动基元：学习运动行为的吸引子模型", "shortTitle": "DMP", "date": "2013", "datePrecision": "year", "year": 2013, "url": "https://doi.org/10.1162/NECO_a_00393", "codeUrl": null, "category": "运动表示", "tags": [ "运动基元", "示范学习", "稳定性" ], "tier": "foundation", "summary": "从稳定弹簧阻尼系统出发，用相位驱动的可学习强迫项编码轨迹形状，分别构造点吸引子和周期吸引子；基函数权重可由示范回归。时间常数、目标和振幅控制执行变化，空间/相位耦合项支持障碍响应、停顿及多关节协调。", "abstractZh": "从稳定弹簧阻尼系统出发，用相位驱动的可学习强迫项编码轨迹形状，分别构造点吸引子和周期吸引子；基函数权重可由示范回归。时间常数、目标和振幅控制执行变化，空间/相位耦合项支持障碍响应、停顿及多关节协调。\n文章系统整理理论并汇集仿真和机器人示例，包括30自由度人形网球/打鼓、Sarcos Slave七自由度手臂放杯与移动目标、障碍耦合、节律同步和动作识别。二维环形轨迹实验专门比较不同坐标系下目标变化造成的形状泛化。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "both", "robots": [ "Sarcos Slave 7-DOF", "30自由度人形机器人（正文未在此命名）" ], "codeStatus": "unknown", "status": "论文提供历史MATLAB下载入口；当前可用性与许可未核实。", "trainingNote": "用示范轨迹拟合强迫项；实际执行仍需底层跟踪控制器。", "whyUseful": "应先用表1方程拟合离散/周期轨迹，检查相位、时间缩放、端点和耦合，再接入低层控制；区分综述所引历史机器人结果与新量化实验。本次已视觉核对图14，未运行历史MATLAB实现。", "contribution": "从稳定弹簧阻尼系统出发，用相位驱动的可学习强迫项编码轨迹形状，分别构造点吸引子和周期吸引子；基函数权重可由示范回归。时间常数、目标和振幅控制执行变化，空间/相位耦合项支持障碍响应、停顿及多关节协调。", "limitations": "稳定性性质针对所定义动力系统及耦合条件，不自动保证机器人力矩可实现、无碰撞或安全。DMP产生运动学计划，需PD/逆动力学跟踪；坐标选择、接近零位移和额外耦合会改变性质，非通用视觉决策器。", "caveats": "按期刊卷期记2013年，不补造日月；部分机器人例子回顾作者早期工作。", "evidence": [ { "url": "https://www.pure.ed.ac.uk/ws/portalfiles/portal/7874487/NECO_a_00393.pdf", "note": "作者机构保存的出版版：Neural Computation 25，328–373，2013；第3节仿真与硬件，图9为Sarcos Slave；脚注2为历史代码入口。" }, { "url": "https://www.pure.ed.ac.uk/ws/portalfiles/portal/7874487/NECO_a_00393.pdf", "note": "§2.1 Table 1; §3.1：弹簧阻尼、强迫项及运动学计划须低层控制器跟踪。" }, { "url": "https://www.pure.ed.ac.uk/ws/portalfiles/portal/7874487/NECO_a_00393.pdf", "note": "§3.2 Figure 9; §3.4 Figure 14：Sarcos放杯例和坐标系导致形状失真；PDF含一页库封面。" } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "基础理论", "模仿学习" ], "robotFilters": [ "Sarcos Slave 7-DOF" ], "verificationNote": "2026-10-04核对出版版全文；只列正文明确命名的硬件。", "original": { "id": "dmp", "originalSourceUrl": "https://www.pure.ed.ac.uk/ws/portalfiles/portal/7874487/NECO_a_00393.pdf", "finalDownloadUrl": "https://www.pure.ed.ac.uk/ws/portalfiles/portal/7874487/NECO_a_00393.pdf", "sha256": "3b7c8d3a141c7edd0253f8056dee39155fe9ed678da75676bb41a39324b534f4", "bytes": 2741402, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:21.171023+00:00", "publicationDate": "2013", "publicationDateSourceUrl": "https://www.pure.ed.ac.uk/ws/portalfiles/portal/7874487/NECO_a_00393.pdf", "license": "Copyright 2013 Massachusetts Institute of Technology; repository says rights retained by authors/copyright owners.", "licenseSourceUrl": "https://www.pure.ed.ac.uk/ws/portalfiles/portal/7874487/NECO_a_00393.pdf", "licenseStatus": "rights_notice_verified", "licenseScope": "paper; does not establish code license", "pages": 47, "sourceIdentityStatus": "title_match_in_pdf", "documentVersion": "Publisher PDF (version of record), with repository cover page", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.548147+00:00" }, "originalAnalysis": { "id": "dmp", "sourceUrl": "https://www.pure.ed.ac.uk/ws/portalfiles/portal/7874487/NECO_a_00393.pdf", "sectionsRead": [ "§2.1", "Table 1", "§3.1–3.4", "Figures 8–9,14", "§5" ], "methodsZh": "从稳定弹簧阻尼系统出发，用相位驱动的可学习强迫项编码轨迹形状，分别构造点吸引子和周期吸引子；基函数权重可由示范回归。时间常数、目标和振幅控制执行变化，空间/相位耦合项支持障碍响应、停顿及多关节协调。", "experimentsZh": "文章系统整理理论并汇集仿真和机器人示例，包括30自由度人形网球/打鼓、Sarcos Slave七自由度手臂放杯与移动目标、障碍耦合、节律同步和动作识别。二维环形轨迹实验专门比较不同坐标系下目标变化造成的形状泛化。", "resultsZh": "作者展示同一放杯运动基元在目标移动及障碍接近时连续调节，无需中止重新生成整段计划。图14表明笛卡尔起终点某轴过近会放大或压塌轨迹环，采用局部坐标改变泛化形态；本论文没有统一跨任务成功率排行榜。", "limitationsZh": "稳定性性质针对所定义动力系统及耦合条件，不自动保证机器人力矩可实现、无碰撞或安全。DMP产生运动学计划，需PD/逆动力学跟踪；坐标选择、接近零位移和额外耦合会改变性质，非通用视觉决策器。", "reproductionZh": "应先用表1方程拟合离散/周期轨迹，检查相位、时间缩放、端点和耦合，再接入低层控制；区分综述所引历史机器人结果与新量化实验。本次已视觉核对图14，未运行历史MATLAB实现。", "experimentType": "both", "robots": [ "Sarcos Slave 7-DOF", "30自由度人形机器人（正文未在此命名）" ], "evidenceNotes": [ { "section/page": "§2.1 Table 1; §3.1", "note": "弹簧阻尼、强迫项及运动学计划须低层控制器跟踪。" }, { "section/page": "§3.2 Figure 9; §3.4 Figure 14", "note": "Sarcos放杯例和坐标系导致形状失真；PDF含一页库封面。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04", "corrections": [] }, "experimentNote": "文章系统整理理论并汇集仿真和机器人示例，包括30自由度人形网球/打鼓、Sarcos Slave七自由度手臂放杯与移动目标、障碍耦合、节律同步和动作识别。二维环形轨迹实验专门比较不同坐标系下目标变化造成的形状泛化。", "analysisVerifiedAt": "2026-10-04" }, { "id": "guided-policy-search", "title": "Guided Policy Search", "titleZh": "引导式策略搜索", "shortTitle": "Guided Policy Search", "date": "2013", "datePrecision": "year", "year": 2013, "url": "https://proceedings.mlr.press/v28/levine13.html", "codeUrl": null, "category": "强化学习", "tags": [ "轨迹优化", "策略搜索", "基础算法" ], "tier": "foundation", "summary": "2013版GPS先用微分动态规划DDP找到随机反馈轨迹分布，从中采样初始化神经网络策略，再把这些引导样本和策略自身经验放进正则化重要性采样目标。优化和采样交替进行，必要时适应引导分布，使高回报局部解推动更通用策…", "abstractZh": "2013版GPS先用微分动态规划DDP找到随机反馈轨迹分布，从中采样初始化神经网络策略，再把这些引导样本和策略自身经验放进正则化重要性采样目标。优化和采样交替进行，必要时适应引导分布，使高回报局部解推动更通用策略搜索。\nMuJoCo平面游泳、单腿跳、行走及三维人形跑步，输入关节角/速度，输出关节力矩，带电机噪声。比较无正则、只预训练不用后续引导、重启分布、普通策略梯度和模仿DDP的DAgger；有50隐单元、500步时域等受控配置。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "sim", "robots": [], "codeStatus": "unknown", "status": "本次未核实2013原始实验的官方代码包。", "trainingNote": "模型辅助轨迹优化产生引导样本，再迭代优化参数化策略。", "whyUseful": "先复现低维游泳/跳跃的DDP引导采样及LBFGS正则目标，检查概率密度和融合采样分布，记录真实交互样本数。保留本文版本边界，不用后续GPS的局部线性学习动态、BADMM或视觉策略流程替代。", "contribution": "2013版GPS先用微分动态规划DDP找到随机反馈轨迹分布，从中采样初始化神经网络策略，再把这些引导样本和策略自身经验放进正则化重要性采样目标。优化和采样交替进行，必要时适应引导分布，使高回报局部解推动更通用策略搜索。", "limitations": "依赖可用可微或可数值求导的动力学、奖励和有效DDP初值；重要性权重仍可能高方差，复杂策略目标非凸。仅状态输入仿真控制，没有本篇视觉端到端PR2或真实机器人结果。", "caveats": "不要将后续GPS、PR2视觉控制论文的硬件和实现当成本篇实验。", "evidence": [ { "url": "https://proceedings.mlr.press/v28/levine13.html", "note": "ICML 2013出版记录与摘要；实验明确为模拟控制任务。" }, { "url": "https://proceedings.mlr.press/v28/levine13.pdf", "note": "方法为DDP引导与正则化重要性采样策略优化。" }, { "url": "https://proceedings.mlr.press/v28/levine13.pdf", "note": "3–5 Algorithm 1：DDP反馈样本、重要性采样和自适应正则。" }, { "url": "https://proceedings.mlr.press/v28/levine13.pdf", "note": "6：模拟形态、消融和适应引导分布效果。" } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "基础理论", "强化学习" ], "robotFilters": [], "verificationNote": "2026-10-04核对PMLR论文；采用明确年份，避免混用会议与电子出版日。", "original": { "id": "guided-policy-search", "originalSourceUrl": "https://proceedings.mlr.press/v28/levine13.pdf", "finalDownloadUrl": "https://proceedings.mlr.press/v28/levine13.pdf", "sha256": "f0ea0f19842973f8d675d6ba6dec5b6bfddc576211dadb0e02dccf452e97790a", "bytes": 1594905, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:28.946864+00:00", "publicationDate": "2013-06", "publicationDateSourceUrl": "https://proceedings.mlr.press/v28/levine13.html", "license": null, "licenseSourceUrl": null, "licenseStatus": "not_verified", "pages": 9, "sourceIdentityStatus": "title_match_in_pdf", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.550562+00:00" }, "originalAnalysis": { "id": "guided-policy-search", "sourceUrl": "https://proceedings.mlr.press/v28/levine13.pdf", "sectionsRead": [ "3", "4.1–4.3", "5", "6", "Algorithm 1; Figs 2–5" ], "methodsZh": "2013版GPS先用微分动态规划DDP找到随机反馈轨迹分布，从中采样初始化神经网络策略，再把这些引导样本和策略自身经验放进正则化重要性采样目标。优化和采样交替进行，必要时适应引导分布，使高回报局部解推动更通用策略搜索。", "experimentsZh": "MuJoCo平面游泳、单腿跳、行走及三维人形跑步，输入关节角/速度，输出关节力矩，带电机噪声。比较无正则、只预训练不用后续引导、重启分布、普通策略梯度和模仿DDP的DAgger；有50隐单元、500步时域等受控配置。", "resultsZh": "完整方法学得多个步态并达到初始引导轨迹回报，正则项与搜索期间持续引导均重要；只利用起始样本可能陷入坏局部解。适应指导分布帮助融合多个初始示范并改善策略，展示优化与学习结合的可行性。", "limitationsZh": "依赖可用可微或可数值求导的动力学、奖励和有效DDP初值；重要性权重仍可能高方差，复杂策略目标非凸。仅状态输入仿真控制，没有本篇视觉端到端PR2或真实机器人结果。", "reproductionZh": "先复现低维游泳/跳跃的DDP引导采样及LBFGS正则目标，检查概率密度和融合采样分布，记录真实交互样本数。保留本文版本边界，不用后续GPS的局部线性学习动态、BADMM或视觉策略流程替代。", "experimentType": "sim", "robots": [], "corrections": [ "应与后续GPS/PR2论文区分，2013原文评估全为MuJoCo状态输入模拟控制。" ], "evidenceNotes": [ { "note": "DDP反馈样本、重要性采样和自适应正则。", "section": "3–5 Algorithm 1" }, { "note": "模拟形态、消融和适应引导分布效果。", "section": "6" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:51:00Z", "sourceVersion": "ICML 2013 PMLR v28 original", "originalSha256": "f0ea0f19842973f8d675d6ba6dec5b6bfddc576211dadb0e02dccf452e97790a" }, "experimentNote": "MuJoCo平面游泳、单腿跳、行走及三维人形跑步，输入关节角/速度，输出关节力矩，带电机噪声。比较无正则、只预训练不用后续引导、重启分布、普通策略梯度和模仿DDP的DAgger；有50隐单元、500步时域等受控配置。", "analysisVerifiedAt": "2026-10-04T13:51:00Z" }, { "id": "arxiv-1011.0686", "title": "A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning", "titleZh": "将模仿学习与结构化预测归约为无悔在线学习", "shortTitle": "DAgger", "date": "2010-11-02", "datePrecision": "day", "year": 2010, "url": "https://arxiv.org/abs/1011.0686", "codeUrl": null, "category": "模仿学习", "tags": [ "分布偏移", "交互式模仿", "基础算法" ], "tier": "foundation", "summary": "DAgger迭代执行当前策略或其与专家的混合，在实际访问状态查询专家动作，将样本加入累计数据集后重新监督训练。重点是改变训练状态分布以覆盖学习者自身错误；混合专家比例逐渐降低，最终可部署单一策略。无悔在线学习把…", "abstractZh": "DAgger迭代执行当前策略或其与专家的混合，在实际访问状态查询专家动作，将样本加入累计数据集后重新监督训练。重点是改变训练状态分布以覆盖学习者自身错误；混合专家比例逐渐降低，最终可部署单一策略。无悔在线学习把训练损失联系到策略诱导分布下的表现。\n原实验含Super Tux Kart赛车、随机生成Super Mario关卡及手写字符序列，没有实体机器人。赛车用人类操作者标转向和5Hz线性回归；Mario用可访问完整状态的规划器当专家、四个线性SVM，每轮5000样本共20轮；手写约6600词、52000字符做十折评测。", "translationType": "中文原文分析（非逐字全文翻译）", "experimentType": "sim", "robots": [], "codeStatus": "unknown", "status": "本次未核实原论文官方代码。", "trainingNote": "交互式监督学习；每轮需要专家标注策略访问的状态。", "whyUseful": "复现应记录每轮累计数据、专家混合概率、专家信息权限、查询量和验证选模。训练时安全接管与专家离线标注是不同环节；应比较等标签预算及重复种子，避免把仅收集专家正常轨迹的行为克隆混称DAgger。", "contribution": "DAgger迭代执行当前策略或其与专家的混合，在实际访问状态查询专家动作，将样本加入累计数据集后重新监督训练。重点是改变训练状态分布以覆盖学习者自身错误；混合专家比例逐渐降低，最终可部署单一策略。无悔在线学习把训练损失联系到策略诱导分布下的表现。", "limitations": "需要可查询的专家为偏离专家轨迹的状态标注，交互和安全成本没有消失。线性时域优势依赖代理损失、无悔及专家恢复代价等条件，不能保证任意不可恢复任务免误差累积；游戏结果也不等于真实自主驾驶能力。", "caveats": "日期为预印本首发；正式会议为AISTATS 2011。原论文不是实体机器人实验。", "evidence": [ { "url": "https://arxiv.org/abs/1011.0686", "note": "标题、首发日期与方法摘要。" }, { "url": "https://arxiv.org/pdf/1011.0686", "note": "第5节：Super Tux Kart、Super Mario Bros.与手写序列标注。" }, { "url": "https://proceedings.mlr.press/v15/ross11a.html", "note": "正式会议记录为2011年。" }, { "url": "https://arxiv.org/pdf/1011.0686v3", "note": "3 算法3.1：Roll outβiexpert+(1−βi)learner;query expert at visited states;aggregate all data;return best validation policy." }, { "url": "https://arxiv.org/pdf/1011.0686v3", "note": "4：No-regret bounded-loss assumptions;cost-to-go factoru affects task-cost bound;finite-sample guarantees stated separately." }, { "url": "https://arxiv.org/pdf/1011.0686v3", "note": "5.1/5.2：Human steering expert in kart;Mario full-state planner expert;20iterations×5000data;3030vs2980distance." }, { "url": "https://arxiv.org/pdf/1011.0686v3", "note": "5.3：10-fold handwritten dataset;85.5%DAgger vs83.6%supervised vs82%unstructured." } ], "verification": "verified", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "directions": [ "基础理论", "模仿学习" ], "robotFilters": [], "verificationNote": "2026-10-04核对论文及PMLR记录；未将后续真机应用归入原论文。", "original": { "id": "arxiv-1011.0686", "originalSourceUrl": "https://arxiv.org/pdf/1011.0686v3", "versionedOriginalUrl": "https://arxiv.org/pdf/1011.0686v3", "finalDownloadUrl": "https://arxiv.org/pdf/1011.0686", "sha256": "12b032cfd451b8ace4b18145dbe24382d2cbe2eda6b562331bc86799bbaafd5e", "bytes": 551229, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:41:57.020639+00:00", "publicationDate": "2010/11/02", "publicationDateSourceUrl": "https://arxiv.org/abs/1011.0686", "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/", "licenseSourceUrl": "https://arxiv.org/abs/1011.0686", "licenseStatus": "license_url_verified", "licenseScope": "paper; does not establish code license", "metadataStatus": "checked", "metadataSourceUrl": "https://arxiv.org/abs/1011.0686", "pages": 9, "sourceIdentityStatus": "title_match", "sourceTitle": "A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning", "sourceVersion": "1011.0686v3", "sourceVersionDate": "2011/03/16", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.553278+00:00", "unversionedOriginalSourceUrl": "https://arxiv.org/pdf/1011.0686" }, "originalAnalysis": { "id": "arxiv-1011.0686", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:47:58.739415+00:00", "sourceUrl": "https://arxiv.org/pdf/1011.0686v3", "sectionsRead": [ "2 分布偏移与代价界", "3 算法3.1数据聚合", "4 在线无悔与有限样本分析", "5.1–5.3 赛车、Mario和手写识别" ], "methodsZh": "DAgger迭代执行当前策略或其与专家的混合，在实际访问状态查询专家动作，将样本加入累计数据集后重新监督训练。重点是改变训练状态分布以覆盖学习者自身错误；混合专家比例逐渐降低，最终可部署单一策略。无悔在线学习把训练损失联系到策略诱导分布下的表现。", "experimentsZh": "原实验含Super Tux Kart赛车、随机生成Super Mario关卡及手写字符序列，没有实体机器人。赛车用人类操作者标转向和5Hz线性回归；Mario用可访问完整状态的规划器当专家、四个线性SVM，每轮5000样本共20轮；手写约6600词、52000字符做十折评测。", "resultsZh": "作者报告赛车15轮后评测未掉出赛道，优于单纯重复专家轨迹及SMILe。Mario按0.5衰减专家混合取得约3030平均行进距离，首轮后不用专家混合约2980，而关卡全长约4200至4300；手写准确率85.5%，普通监督83.6%、无结构82%。", "limitationsZh": "需要可查询的专家为偏离专家轨迹的状态标注，交互和安全成本没有消失。线性时域优势依赖代理损失、无悔及专家恢复代价等条件，不能保证任意不可恢复任务免误差累积；游戏结果也不等于真实自主驾驶能力。", "reproductionZh": "复现应记录每轮累计数据、专家混合概率、专家信息权限、查询量和验证选模。训练时安全接管与专家离线标注是不同环节；应比较等标签预算及重复种子，避免把仅收集专家正常轨迹的行为克隆混称DAgger。", "evidenceNotes": [ { "section/page": "3 算法3.1", "note": "Roll outβiexpert+(1−βi)learner;query expert at visited states;aggregate all data;return best validation policy." }, { "section/page": "4", "note": "No-regret bounded-loss assumptions;cost-to-go factoru affects task-cost bound;finite-sample guarantees stated separately." }, { "section/page": "5.1/5.2", "note": "Human steering expert in kart;Mario full-state planner expert;20iterations×5000data;3030vs2980distance." }, { "section/page": "5.3", "note": "10-fold handwritten dataset;85.5%DAgger vs83.6%supervised vs82%unstructured." } ], "experimentType": "sim", "robots": [], "corrections": [ "原实验含游戏仿真与手写序列数据，无实体机器人。" ], "sourceVersion": "1011.0686v3", "originalSha256": "12b032cfd451b8ace4b18145dbe24382d2cbe2eda6b562331bc86799bbaafd5e" }, "experimentNote": "原实验含Super Tux Kart赛车、随机生成Super Mario关卡及手写字符序列，没有实体机器人。赛车用人类操作者标转向和5Hz线性回归；Mario用可访问完整状态的规划器当专家、四个线性SVM，每轮5000样本共20轮；手写约6600词、52000字符做十折评测。", "analysisVerifiedAt": "2026-10-04T13:47:58.739415+00:00" }, { "id": "monoslam-2007", "title": "MonoSLAM: Real-Time Single Camera SLAM", "titleZh": "MonoSLAM：实时单目相机同步定位与建图", "date": "2007", "datePrecision": "year", "year": 2007, "url": "https://robots.ox.ac.uk/ActiveVision/Publications/davison_etal_pami2007/davison_etal_pami2007.html", "paperUrl": "https://robots.ox.ac.uk/ActiveVision/Publications/davison_etal_pami2007/davison_etal_pami2007.html", "codeUrl": "https://www.doc.ic.ac.uk/~ajd/Scene/", "summary": "MonoSLAM用全协方差EKF联合估计相机位姿、速度和稀疏地标，以平滑运动预测压缩特征搜索范围；图像块作为长期地标，处理单目深度初始化与视角形变。地图保存相关不确定性，使重访时整体位置同步修正。", "abstractZh": "MonoSLAM用全协方差EKF联合估计相机位姿、速度和稀疏地标，以平滑运动预测压缩特征搜索范围；图像块作为长期地标，处理单目深度初始化与视角形变。地图保存相关不确定性，使重访时整体位置同步修正。\n展示30Hz手持增强现实与HRP-2室内走半径0.75米圆周，约30秒运动中分五段暂停。人形版本额外使用胸部陀螺仪；单相机约90度视场，并以测得位置的自然/人工特征初始化。另用带铅垂线的手持相机检验桌面四角真值。", "category": "视觉 SLAM / 概率估计", "tags": [ "单目 SLAM", "EKF", "主动视觉", "稀疏建图" ], "directions": [ "导航与建图" ], "tier": "foundation", "experimentType": "real", "robots": [ "HRP-2" ], "robotFilters": [ "HRP-2" ], "codeStatus": "open", "status": "实现已开源", "trainingNote": "几何估计系统，无神经网络训练。作者公开 SceneLib 1.0 与 MonoSLAMGlow；官网明确该版本并未实现全部后续技术，例如 patch warping，不能当作 2007 论文全部功能的逐项复现。", "contribution": "MonoSLAM用全协方差EKF联合估计相机位姿、速度和稀疏地标，以平滑运动预测压缩特征搜索范围；图像块作为长期地标，处理单目深度初始化与视角形变。地图保存相关不确定性，使重访时整体位置同步修正。", "whyUseful": "复现应区分标准初始化靶与HRP-2测量特征，保留相机径向标定、噪声、陀螺更新和特征管理。SceneLib为参考入口，发布版本未必包含论文全部后续patch功能；本轮未编译，也未复现机器人轨迹。", "limitations": "EKF存储/计算随地标平方增长，文中实时约100特征；局限小房间、足够纹理和较平滑运动。实机并非纯视觉，初始化有已知几何；局部重访不漂移不能推广到无限规模或任意遮挡/急运动。", "license": "LGPL（SceneLib 官网声明；具体版本未核实）", "evidence": [ { "url": "https://spiral.imperial.ac.uk/entities/publication/5ceb6f60-d4b1-4d1e-a1bf-b518e74ff2c9", "note": "作者机构元数据确认 2007 年期刊论文，月份为 6 月；未补造具体日。" }, { "url": "https://www.robots.ox.ac.uk/~lav/Papers/davison_etal_pami2007/davison_etal_pami2007.pdf", "note": "全文 §V 核实 HRP-2、单目相机和陀螺仪融合实验。" }, { "url": "https://www.doc.ic.ac.uk/~ajd/Scene/", "note": "作者源码页确认 LGPL、MonoSLAMGlow 与未实现 patch warping 的发布范围。" }, { "url": "https://www.robots.ox.ac.uk/~lav/Papers/davison_etal_pami2007/davison_etal_pami2007.pdf", "note": "§V-A–B：额外宽角单相机、已知位置特征初始化并融合陀螺仪。" }, { "url": "https://www.robots.ox.ac.uk/~lav/Papers/davison_etal_pami2007/davison_etal_pami2007.pdf", "note": "§V-C/Fig.10：半径0.75米、约30秒、暂停五段的人形闭环实验。" }, { "url": "https://www.robots.ox.ac.uk/~lav/Papers/davison_etal_pami2007/davison_etal_pami2007.pdf", "note": "§VI-A–B：桌面四点量测，厘米级偏差与约19ms典型处理。" }, { "url": "https://www.robots.ox.ac.uk/~lav/Papers/davison_etal_pami2007/davison_etal_pami2007.pdf", "note": "Author manuscript p.19, Fig.11 and table (visual inspection)：已核对桌面铅垂线真值测试和四点统计；不是HRP-2轨迹精度表。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "把递推概率 SLAM 落到实时单目视觉，是理解后续关键帧优化路线的重要起点。", "robotNote": "HRP-2 为实机；另有手持相机测试。", "experimentNote": "展示30Hz手持增强现实与HRP-2室内走半径0.75米圆周，约30秒运动中分五段暂停。人形版本额外使用胸部陀螺仪；单相机约90度视场，并以测得位置的自然/人工特征初始化。另用带铅垂线的手持相机检验桌面四角真值。", "projectUrl": null, "shortTitle": "", "original": { "id": "monoslam-2007", "originalSourceUrl": "https://www.robots.ox.ac.uk/~lav/Papers/davison_etal_pami2007/davison_etal_pami2007.pdf", "finalDownloadUrl": "https://www.robots.ox.ac.uk/~lav/Papers/davison_etal_pami2007/davison_etal_pami2007.pdf", "sha256": "a564a4fd07e9a9c52bf0dcf46c30e691367df89b78d351200e16b29508a91e30", "bytes": 943903, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:37.060161+00:00", "publicationDate": "2007", "publicationDateSourceUrl": "https://www.robots.ox.ac.uk/~lav/Papers/davison_etal_pami2007/davison_etal_pami2007.html", "license": "Copyright retained by authors or copyright holders; reposting usually requires permission.", "licenseSourceUrl": "https://www.robots.ox.ac.uk/~lav/Papers/davison_etal_pami2007/davison_etal_pami2007.html", "licenseStatus": "rights_notice_verified", "licenseScope": "paper; does not establish code license", "pages": 24, "sourceIdentityStatus": "title_match_in_pdf", "documentVersion": "Author-hosted manuscript (24 pages)", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.555067+00:00" }, "originalAnalysis": { "id": "monoslam-2007", "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04T13:50:35.090157+00:00", "sourceUrl": "https://www.robots.ox.ac.uk/~lav/Papers/davison_etal_pami2007/davison_etal_pami2007.pdf", "sectionsRead": [ "§III", "§IV", "§V-A–C", "§VI-A–C", "Figs.9–11", "§VII" ], "methodsZh": "MonoSLAM用全协方差EKF联合估计相机位姿、速度和稀疏地标，以平滑运动预测压缩特征搜索范围；图像块作为长期地标，处理单目深度初始化与视角形变。地图保存相关不确定性，使重访时整体位置同步修正。", "experimentsZh": "展示30Hz手持增强现实与HRP-2室内走半径0.75米圆周，约30秒运动中分五段暂停。人形版本额外使用胸部陀螺仪；单相机约90度视场，并以测得位置的自然/人工特征初始化。另用带铅垂线的手持相机检验桌面四角真值。", "resultsZh": "作者桌面实验定位通常厘米级、抖动约1–2厘米，但部分坐标存在更大系统偏差。HRP-2闭环后不确定性缩小；1.6GHz Pentium M典型整帧处理约19毫秒，支持该规模30Hz。", "limitationsZh": "EKF存储/计算随地标平方增长，文中实时约100特征；局限小房间、足够纹理和较平滑运动。实机并非纯视觉，初始化有已知几何；局部重访不漂移不能推广到无限规模或任意遮挡/急运动。", "reproductionZh": "复现应区分标准初始化靶与HRP-2测量特征，保留相机径向标定、噪声、陀螺更新和特征管理。SceneLib为参考入口，发布版本未必包含论文全部后续patch功能；本轮未编译，也未复现机器人轨迹。", "experimentType": "real", "robots": [ "HRP-2" ], "evidenceNotes": [ { "section/page": "§V-A–B", "note": "额外宽角单相机、已知位置特征初始化并融合陀螺仪。" }, { "section/page": "§V-C/Fig.10", "note": "半径0.75米、约30秒、暂停五段的人形闭环实验。" }, { "section/page": "§VI-A–B", "note": "桌面四点量测，厘米级偏差与约19ms典型处理。" }, { "section/page": "Author manuscript p.19, Fig.11 and table (visual inspection)", "note": "已核对桌面铅垂线真值测试和四点统计；不是HRP-2轨迹精度表。" } ] }, "analysisVerifiedAt": "2026-10-04T13:50:35.090157+00:00" }, { "id": "ptam-2007", "title": "Parallel Tracking and Mapping for Small AR Workspaces", "titleZh": "PTAM：面向小型增强现实工作空间的并行跟踪与建图", "date": "2007", "datePrecision": "year", "year": 2007, "url": "https://www.robots.ox.ac.uk/~lav/Papers/klein_murray_ismar2007/klein_murray_ismar2007.html", "paperUrl": "https://www.robots.ox.ac.uk/~lav/Papers/klein_murray_ismar2007/klein_murray_ismar2007.html", "codeUrl": "https://github.com/Oxford-PTAM/PTAM-GPL", "summary": "把相机实时跟踪与关键帧地图优化分到并行线程：跟踪端粗到细匹配大量FAST局部纹理，用鲁棒重投影估计位姿；建图端三角化新点并局部/全局束调整。用户平移相机采两关键帧完成五点初始化，避免每帧共同滤波全部地图。", "abstractZh": "把相机实时跟踪与关键帧地图优化分到并行线程：跟踪端粗到细匹配大量FAST局部纹理，用鲁棒重投影估计位姿；建图端三角化新点并局部/全局束调整。用户平移相机采两关键帧完成五点初始化，避免每帧共同滤波全部地图。\n主要用真实手持相机在桌面/办公室做AR定位，另以合成序列与EKF-SLAM比较。典型1656帧桌面序列形成57关键帧及4997点；在2.66GHz双核PC上测跟踪、局部和全局优化随地图规模变化。", "category": "视觉 SLAM / 关键帧优化", "tags": [ "PTAM", "关键帧", "Bundle Adjustment", "增强现实" ], "directions": [ "导航与建图" ], "tier": "foundation", "experimentType": "both", "robots": [], "robotFilters": [], "codeStatus": "open", "status": "实现已开源", "trainingNote": "无需策略训练；官方 GPL 重发布仓库提供 C++ 实现，基于历史 PTAM v1.0-r114。编译依赖较旧，初始化需要相机平移。", "contribution": "把相机实时跟踪与关键帧地图优化分到并行线程：跟踪端粗到细匹配大量FAST局部纹理，用鲁棒重投影估计位姿；建图端三角化新点并局部/全局束调整。用户平移相机采两关键帧完成五点初始化，避免每帧共同滤波全部地图。", "whyUseful": "需匹配相机标定、图像金字塔、patch搜索阈值和异步调度，分别测跟踪与建图延迟。传统系统无需策略训练；原论文与后续公开代码新增模块应分开说明，本次未编译运行。", "limitations": "针对小工作区，相机需纹理充分、场景较静态且遮挡有限；模糊、重复纹理、错误关键帧可损坏地图。不是完整大回环或自主导航系统；初始化依用户平移和按键，纯旋转不足以三角化。", "license": "GPL-3.0（PTAM-GPL 重发布版本）", "evidence": [ { "url": "https://www.robots.ox.ac.uk/~lav/Papers/klein_murray_ismar2007/klein_murray_ismar2007.html", "note": "牛津作者页确认 ISMAR 2007、并行线程和小型手持 AR 场景。" }, { "url": "https://www.robots.ox.ac.uk/~gk/PTAM/", "note": "作者项目页说明代码对应版本、初始化平移要求及 GPL 重发布。" }, { "url": "https://github.com/Oxford-PTAM/PTAM-GPL", "note": "官方重发布仓库确认 GPLv3 和 v1.0-r114 源码血缘。" }, { "url": "https://www.robots.ox.ac.uk/~lav/Papers/klein_murray_ismar2007/klein_murray_ismar2007.pdf", "note": "§2,5–6：并行线程、关键帧BA和需用户合作的两帧初始化。" }, { "url": "https://www.robots.ox.ac.uk/~lav/Papers/klein_murray_ismar2007/klein_murray_ismar2007.pdf", "note": "§7 Tables 1–2; §8：19.2毫秒跟踪、合成比较、扩图规模与大回环限制。" } ], "verification": "verified", "translationType": "中文原文分析（非逐字全文翻译）", "fullTextTranslation": "未提供", "verifiedAt": "2026-10-04", "verificationNote": "核对所列一手论文、项目或代码来源；已核对的范围见证据说明，未独立复现实验或执行代码。", "timelineNote": "将跟踪与建图并行化，形成关键帧优化型视觉 SLAM 的代表性架构。", "experimentNote": "主要用真实手持相机在桌面/办公室做AR定位，另以合成序列与EKF-SLAM比较。典型1656帧桌面序列形成57关键帧及4997点；在2.66GHz双核PC上测跟踪、局部和全局优化随地图规模变化。", "robotNote": "论文针对手持相机，未添加机器人型号。", "projectUrl": null, "shortTitle": "", "original": { "id": "ptam-2007", "originalSourceUrl": "https://www.robots.ox.ac.uk/~lav/Papers/klein_murray_ismar2007/klein_murray_ismar2007.pdf", "finalDownloadUrl": "https://www.robots.ox.ac.uk/~lav/Papers/klein_murray_ismar2007/klein_murray_ismar2007.pdf", "sha256": "a71a700a0a56b3893915c43f78b741d47fdbe60e1ff58071a11d6881b9b0065f", "bytes": 1618643, "downloadStatus": "downloaded", "retrievedAt": "2026-10-04T13:35:45.509195+00:00", "publicationDate": "2007", "publicationDateSourceUrl": "https://www.robots.ox.ac.uk/~lav/Papers/klein_murray_ismar2007/klein_murray_ismar2007.html", "license": "Copyright retained by authors or copyright holders; reposting usually requires permission.", "licenseSourceUrl": "https://www.robots.ox.ac.uk/~lav/Papers/klein_murray_ismar2007/klein_murray_ismar2007.html", "licenseStatus": "rights_notice_verified", "licenseScope": "paper; does not establish code license", "pages": 10, "sourceIdentityStatus": "title_match_in_pdf", "archiveValidationStatus": "validated", "archiveValidatedAt": "2026-10-04T13:45:15.557373+00:00" }, "originalAnalysis": { "id": "ptam-2007", "sourceUrl": "https://www.robots.ox.ac.uk/~lav/Papers/klein_murray_ismar2007/klein_murray_ismar2007.pdf", "sectionsRead": [ "§2,4–6", "§7.1–7.4", "Tables 1–2", "§8–9" ], "methodsZh": "把相机实时跟踪与关键帧地图优化分到并行线程：跟踪端粗到细匹配大量FAST局部纹理，用鲁棒重投影估计位姿；建图端三角化新点并局部/全局束调整。用户平移相机采两关键帧完成五点初始化，避免每帧共同滤波全部地图。", "experimentsZh": "主要用真实手持相机在桌面/办公室做AR定位，另以合成序列与EKF-SLAM比较。典型1656帧桌面序列形成57关键帧及4997点；在2.66GHz双核PC上测跟踪、局部和全局优化随地图规模变化。", "resultsZh": "作者报告4000点地图典型跟踪19.2毫秒，实用规模约6000点/150关键帧。100–149关键帧全局束调整平均6.9秒，由后台执行；实时的是跟踪而非所有优化。最多11000点/280关键帧地图已影响继续扩图。", "limitationsZh": "针对小工作区，相机需纹理充分、场景较静态且遮挡有限；模糊、重复纹理、错误关键帧可损坏地图。不是完整大回环或自主导航系统；初始化依用户平移和按键，纯旋转不足以三角化。", "reproductionZh": "需匹配相机标定、图像金字塔、patch搜索阈值和异步调度，分别测跟踪与建图延迟。传统系统无需策略训练；原论文与后续公开代码新增模块应分开说明，本次未编译运行。", "experimentType": "both", "robots": [], "evidenceNotes": [ { "section/page": "§2,5–6", "note": "并行线程、关键帧BA和需用户合作的两帧初始化。" }, { "section/page": "§7 Tables 1–2; §8", "note": "19.2毫秒跟踪、合成比较、扩图规模与大回环限制。" } ], "corrections": [ { "field": "experimentType", "value": "both", "reason": "真实手持相机为主，§7.3另有合成序列比较；不是机器人硬件实验。" } ], "analysisStatus": "full_text_sections", "analyzedAt": "2026-10-04" }, "analysisVerifiedAt": "2026-10-04" } ], "revision": "2026-10-04-library-v7" }